This is the multi-page printable view of this section. Click here to print.

Return to the regular view of this page.

Administer a Cluster

Learn common tasks for administering a cluster.

1 - Administration with kubeadm

If you don't yet have a cluster, visit bootstrapping clusters with kubeadm.

The tasks in this section are aimed at people administering an existing cluster:

1.1 - Adding Linux worker nodes

This page explains how to add Linux worker nodes to a kubeadm cluster.

Before you begin

Adding Linux worker nodes

To add new Linux worker nodes to your cluster do the following for each machine:

  1. Connect to the machine by using SSH or another method.
  2. Run the command that was output by kubeadm init. For example:
sudo kubeadm join --token <token> <control-plane-host>:<control-plane-port> --discovery-token-ca-cert-hash sha256:<hash>

Additional information for kubeadm join

If you do not have the token, you can get it by running the following command on the control plane node:

# Run this on a control plane node
sudo kubeadm token list

The output is similar to this:

TOKEN                    TTL  EXPIRES              USAGES           DESCRIPTION            EXTRA GROUPS
8ewj1p.9r9hcjoqgajrj4gi  23h  2018-06-12T02:51:28Z authentication,  The default bootstrap  system:
                                                   signing          token generated by     bootstrappers:
                                                                    'kubeadm init'.        kubeadm:
                                                                                           default-node-token

By default, node join tokens expire after 24 hours. If you are joining a node to the cluster after the current token has expired, you can create a new token by running the following command on the control plane node:

# Run this on a control plane node
sudo kubeadm token create

The output is similar to this:

5didvk.d09sbcov8ph2amjw

If you don't have the value of --discovery-token-ca-cert-hash, you can get it by running the following commands on the control plane node:

# Run this on a control plane node
sudo cat /etc/kubernetes/pki/ca.crt | openssl x509 -pubkey  | openssl rsa -pubin -outform der 2>/dev/null | \
   openssl dgst -sha256 -hex | sed 's/^.* //'

The output is similar to:

8cb2de97839780a412b93877f8507ad6c94f73add17d5d7058e91741c9d5ec78

The output of the kubeadm join command should look something like:

[preflight] Running pre-flight checks

... (log output of join workflow) ...

Node join complete:
* Certificate signing request sent to control-plane and response
  received.
* Kubelet informed of new secure connection details.

Run 'kubectl get nodes' on control-plane to see this machine join.

A few seconds later, you should notice this node in the output from kubectl get nodes. (for example, run kubectl on a control plane node).

What's next

1.2 - Adding Windows worker nodes

FEATURE STATE: Kubernetes v1.18 [beta]

This page explains how to add Windows worker nodes to a kubeadm cluster.

Before you begin

Adding Windows worker nodes

Do the following for each machine:

  1. Open a PowerShell session on the machine.
  2. Make sure you are Administrator or a privileged user.

Then proceed with the steps outlined below.

Install containerd

To install containerd, first run the following command:

curl.exe -LO https://raw.githubusercontent.com/kubernetes-sigs/sig-windows-tools/master/hostprocess/Install-Containerd.ps1

Then run the following command, but first replace CONTAINERD_VERSION with a recent release from the containerd repository. The version must not have a v prefix. For example, use 1.7.22 instead of v1.7.22:

.\Install-Containerd.ps1 -ContainerDVersion CONTAINERD_VERSION
  • Adjust any other parameters for Install-Containerd.ps1 such as netAdapterName as you need them.
  • Set skipHypervisorSupportCheck if your machine does not support Hyper-V and cannot host Hyper-V isolated containers.
  • If you change the Install-Containerd.ps1 optional parameters CNIBinPath and/or CNIConfigPath you will need to configure the installed Windows CNI plugin with matching values.

Install kubeadm and kubelet

Run the following commands to install kubeadm and the kubelet:

curl.exe -LO https://raw.githubusercontent.com/kubernetes-sigs/sig-windows-tools/master/hostprocess/PrepareNode.ps1
.\PrepareNode.ps1 -KubernetesVersion v1.32
  • Adjust the parameter KubernetesVersion of PrepareNode.ps1 if needed.

Run kubeadm join

Run the command that was output by kubeadm init. For example:

kubeadm join --token <token> <control-plane-host>:<control-plane-port> --discovery-token-ca-cert-hash sha256:<hash>

Additional information about kubeadm join

If you do not have the token, you can get it by running the following command on the control plane node:

# Run this on a control plane node
sudo kubeadm token list

The output is similar to this:

TOKEN                    TTL  EXPIRES              USAGES           DESCRIPTION            EXTRA GROUPS
8ewj1p.9r9hcjoqgajrj4gi  23h  2018-06-12T02:51:28Z authentication,  The default bootstrap  system:
                                                   signing          token generated by     bootstrappers:
                                                                    'kubeadm init'.        kubeadm:
                                                                                           default-node-token

By default, node join tokens expire after 24 hours. If you are joining a node to the cluster after the current token has expired, you can create a new token by running the following command on the control plane node:

# Run this on a control plane node
sudo kubeadm token create

The output is similar to this:

5didvk.d09sbcov8ph2amjw

If you don't have the value of --discovery-token-ca-cert-hash, you can get it by running the following commands on the control plane node:

sudo cat /etc/kubernetes/pki/ca.crt | openssl x509 -pubkey  | openssl rsa -pubin -outform der 2>/dev/null | \
   openssl dgst -sha256 -hex | sed 's/^.* //'

The output is similar to:

8cb2de97839780a412b93877f8507ad6c94f73add17d5d7058e91741c9d5ec78

The output of the kubeadm join command should look something like:

[preflight] Running pre-flight checks

... (log output of join workflow) ...

Node join complete:
* Certificate signing request sent to control-plane and response
  received.
* Kubelet informed of new secure connection details.

Run 'kubectl get nodes' on control-plane to see this machine join.

A few seconds later, you should notice this node in the output from kubectl get nodes. (for example, run kubectl on a control plane node).

Network configuration

CNI setup on clusters mixed with Linux and Windows nodes requires more steps than just running kubectl apply on a manifest file. Additionally, the CNI plugin running on control plane nodes must be prepared to support the CNI plugin running on Windows worker nodes.

Only a few CNI plugins currently support Windows. Below you can find individual setup instructions for them:

Install kubectl for Windows (optional)

See Install and Set Up kubectl on Windows.

What's next

1.3 - Upgrading kubeadm clusters

This page explains how to upgrade a Kubernetes cluster created with kubeadm from version 1.31.x to version 1.32.x, and from version 1.32.x to 1.32.y (where y > x). Skipping MINOR versions when upgrading is unsupported. For more details, please visit Version Skew Policy.

To see information about upgrading clusters created using older versions of kubeadm, please refer to following pages instead:

The Kubernetes project recommends upgrading to the latest patch releases promptly, and to ensure that you are running a supported minor release of Kubernetes. Following this recommendation helps you to to stay secure.

The upgrade workflow at high level is the following:

  1. Upgrade a primary control plane node.
  2. Upgrade additional control plane nodes.
  3. Upgrade worker nodes.

Before you begin

  • Make sure you read the release notes carefully.
  • The cluster should use a static control plane and etcd pods or external etcd.
  • Make sure to back up any important components, such as app-level state stored in a database. kubeadm upgrade does not touch your workloads, only components internal to Kubernetes, but backups are always a best practice.
  • Swap must be disabled.

Additional information

  • The instructions below outline when to drain each node during the upgrade process. If you are performing a minor version upgrade for any kubelet, you must first drain the node (or nodes) that you are upgrading. In the case of control plane nodes, they could be running CoreDNS Pods or other critical workloads. For more information see Draining nodes.
  • The Kubernetes project recommends that you match your kubelet and kubeadm versions. You can instead use a version of kubelet that is older than kubeadm, provided it is within the range of supported versions. For more details, please visit kubeadm's skew against the kubelet.
  • All containers are restarted after upgrade, because the container spec hash value is changed.
  • To verify that the kubelet service has successfully restarted after the kubelet has been upgraded, you can execute systemctl status kubelet or view the service logs with journalctl -xeu kubelet.
  • kubeadm upgrade supports --config with a UpgradeConfiguration API type which can be used to configure the upgrade process.
  • kubeadm upgrade does not support reconfiguration of an existing cluster. Follow the steps in Reconfiguring a kubeadm cluster instead.

Considerations when upgrading etcd

Because the kube-apiserver static pod is running at all times (even if you have drained the node), when you perform a kubeadm upgrade which includes an etcd upgrade, in-flight requests to the server will stall while the new etcd static pod is restarting. As a workaround, it is possible to actively stop the kube-apiserver process a few seconds before starting the kubeadm upgrade apply command. This permits to complete in-flight requests and close existing connections, and minimizes the consequence of the etcd downtime. This can be done as follows on control plane nodes:

killall -s SIGTERM kube-apiserver # trigger a graceful kube-apiserver shutdown
sleep 20 # wait a little bit to permit completing in-flight requests
kubeadm upgrade ... # execute a kubeadm upgrade command

Changing the package repository

If you're using the community-owned package repositories (pkgs.k8s.io), you need to enable the package repository for the desired Kubernetes minor release. This is explained in Changing the Kubernetes package repository document.

Determine which version to upgrade to

Find the latest patch release for Kubernetes 1.32 using the OS package manager:

# Find the latest 1.32 version in the list.
# It should look like 1.32.x-*, where x is the latest patch.
sudo apt update
sudo apt-cache madison kubeadm


# Find the latest 1.32 version in the list.
# It should look like 1.32.x-*, where x is the latest patch.
sudo yum list --showduplicates kubeadm --disableexcludes=kubernetes

Upgrading control plane nodes

The upgrade procedure on control plane nodes should be executed one node at a time. Pick a control plane node that you wish to upgrade first. It must have the /etc/kubernetes/admin.conf file.

Call "kubeadm upgrade"

For the first control plane node

  1. Upgrade kubeadm:

    # replace x in 1.32.x-* with the latest patch version
sudo apt-mark unhold kubeadm && \
sudo apt-get update && sudo apt-get install -y kubeadm='1.32.x-*' && \
sudo apt-mark hold kubeadm


    # replace x in 1.32.x-* with the latest patch version
sudo yum install -y kubeadm-'1.32.x-*' --disableexcludes=kubernetes

  2. Verify that the download works and has the expected version:

    kubeadm version

  3. Verify the upgrade plan:

    sudo kubeadm upgrade plan

    This command checks that your cluster can be upgraded, and fetches the versions you can upgrade to. It also shows a table with the component config version states.

  4. Choose a version to upgrade to, and run the appropriate command. For example:

    # replace x with the patch version you picked for this upgrade
sudo kubeadm upgrade apply v1.32.x

    Once the command finishes you should see:

    [upgrade/successful] SUCCESS! Your cluster was upgraded to "v1.32.x". Enjoy!

[upgrade/kubelet] Now that your control plane is upgraded, please proceed with upgrading your kubelets if you haven't already done so.

  5. Manually upgrade your CNI provider plugin.

    Your Container Network Interface (CNI) provider may have its own upgrade instructions to follow. Check the addons page to find your CNI provider and see whether additional upgrade steps are required.

    This step is not required on additional control plane nodes if the CNI provider runs as a DaemonSet.

For the other control plane nodes

Same as the first control plane node but use:

sudo kubeadm upgrade node

instead of:

sudo kubeadm upgrade apply

Also calling kubeadm upgrade plan and upgrading the CNI provider plugin is no longer needed.

Drain the node

Prepare the node for maintenance by marking it unschedulable and evicting the workloads:

# replace <node-to-drain> with the name of your node you are draining
kubectl drain <node-to-drain> --ignore-daemonsets

Upgrade kubelet and kubectl

  1. Upgrade the kubelet and kubectl:

    # replace x in 1.32.x-* with the latest patch version
sudo apt-mark unhold kubelet kubectl && \
sudo apt-get update && sudo apt-get install -y kubelet='1.32.x-*' kubectl='1.32.x-*' && \
sudo apt-mark hold kubelet kubectl


    # replace x in 1.32.x-* with the latest patch version
sudo yum install -y kubelet-'1.32.x-*' kubectl-'1.32.x-*' --disableexcludes=kubernetes

  2. Restart the kubelet:

    sudo systemctl daemon-reload
sudo systemctl restart kubelet

Uncordon the node

Bring the node back online by marking it schedulable:

# replace <node-to-uncordon> with the name of your node
kubectl uncordon <node-to-uncordon>

Upgrade worker nodes

The upgrade procedure on worker nodes should be executed one node at a time or few nodes at a time, without compromising the minimum required capacity for running your workloads.

The following pages show how to upgrade Linux and Windows worker nodes:

Verify the status of the cluster

After the kubelet is upgraded on all nodes verify that all nodes are available again by running the following command from anywhere kubectl can access the cluster:

kubectl get nodes

The STATUS column should show Ready for all your nodes, and the version number should be updated.

Recovering from a failure state

If kubeadm upgrade fails and does not roll back, for example because of an unexpected shutdown during execution, you can run kubeadm upgrade again. This command is idempotent and eventually makes sure that the actual state is the desired state you declare.

To recover from a bad state, you can also run sudo kubeadm upgrade apply --force without changing the version that your cluster is running.

During upgrade kubeadm writes the following backup folders under /etc/kubernetes/tmp:

  • kubeadm-backup-etcd-<date>-<time>
  • kubeadm-backup-manifests-<date>-<time>

kubeadm-backup-etcd contains a backup of the local etcd member data for this control plane Node. In case of an etcd upgrade failure and if the automatic rollback does not work, the contents of this folder can be manually restored in /var/lib/etcd. In case external etcd is used this backup folder will be empty.

kubeadm-backup-manifests contains a backup of the static Pod manifest files for this control plane Node. In case of a upgrade failure and if the automatic rollback does not work, the contents of this folder can be manually restored in /etc/kubernetes/manifests. If for some reason there is no difference between a pre-upgrade and post-upgrade manifest file for a certain component, a backup file for it will not be written.

How it works

kubeadm upgrade apply does the following:

  • Checks that your cluster is in an upgradeable state:
    • The API server is reachable
    • All nodes are in the Ready state
    • The control plane is healthy
  • Enforces the version skew policies.
  • Makes sure the control plane images are available or available to pull to the machine.
  • Generates replacements and/or uses user supplied overwrites if component configs require version upgrades.
  • Upgrades the control plane components or rollbacks if any of them fails to come up.
  • Applies the new CoreDNS and kube-proxy manifests and makes sure that all necessary RBAC rules are created.
  • Creates new certificate and key files of the API server and backs up old files if they're about to expire in 180 days.

kubeadm upgrade node does the following on additional control plane nodes:

  • Fetches the kubeadm ClusterConfiguration from the cluster.
  • Optionally backups the kube-apiserver certificate.
  • Upgrades the static Pod manifests for the control plane components.
  • Upgrades the kubelet configuration for this node.

kubeadm upgrade node does the following on worker nodes:

  • Fetches the kubeadm ClusterConfiguration from the cluster.
  • Upgrades the kubelet configuration for this node.

1.4 - Upgrading Linux nodes

This page explains how to upgrade a Linux Worker Nodes created with kubeadm.

Before you begin

You need to have shell access to all the nodes, and the kubectl command-line tool must be configured to communicate with your cluster. It is recommended to run this tutorial on a cluster with at least two nodes that are not acting as control plane hosts.

To check the version, enter kubectl version.

Changing the package repository

If you're using the community-owned package repositories (pkgs.k8s.io), you need to enable the package repository for the desired Kubernetes minor release. This is explained in Changing the Kubernetes package repository document.

Upgrading worker nodes

Upgrade kubeadm

Upgrade kubeadm:

# replace x in 1.32.x-* with the latest patch version
sudo apt-mark unhold kubeadm && \
sudo apt-get update && sudo apt-get install -y kubeadm='1.32.x-*' && \
sudo apt-mark hold kubeadm


# replace x in 1.32.x-* with the latest patch version
sudo yum install -y kubeadm-'1.32.x-*' --disableexcludes=kubernetes

Call "kubeadm upgrade"

For worker nodes this upgrades the local kubelet configuration:

sudo kubeadm upgrade node

Drain the node

Prepare the node for maintenance by marking it unschedulable and evicting the workloads:

# execute this command on a control plane node
# replace <node-to-drain> with the name of your node you are draining
kubectl drain <node-to-drain> --ignore-daemonsets

Upgrade kubelet and kubectl

  1. Upgrade the kubelet and kubectl:

    # replace x in 1.32.x-* with the latest patch version
sudo apt-mark unhold kubelet kubectl && \
sudo apt-get update && sudo apt-get install -y kubelet='1.32.x-*' kubectl='1.32.x-*' && \
sudo apt-mark hold kubelet kubectl


    # replace x in 1.32.x-* with the latest patch version
sudo yum install -y kubelet-'1.32.x-*' kubectl-'1.32.x-*' --disableexcludes=kubernetes

  2. Restart the kubelet:

    sudo systemctl daemon-reload
sudo systemctl restart kubelet

Uncordon the node

Bring the node back online by marking it schedulable:

# execute this command on a control plane node
# replace <node-to-uncordon> with the name of your node
kubectl uncordon <node-to-uncordon>

What's next

1.5 - Upgrading Windows nodes

FEATURE STATE: Kubernetes v1.18 [beta]

This page explains how to upgrade a Windows node created with kubeadm.

Before you begin

You need to have shell access to all the nodes, and the kubectl command-line tool must be configured to communicate with your cluster. It is recommended to run this tutorial on a cluster with at least two nodes that are not acting as control plane hosts.

Your Kubernetes server must be at or later than version 1.17. To check the version, enter kubectl version.

Upgrading worker nodes

Upgrade kubeadm

  1. From the Windows node, upgrade kubeadm:

    # replace 1.32.0 with your desired version
curl.exe -Lo <path-to-kubeadm.exe>  "https://dl.k8s.io/v1.32.0/bin/windows/amd64/kubeadm.exe"

Drain the node

  1. From a machine with access to the Kubernetes API, prepare the node for maintenance by marking it unschedulable and evicting the workloads:

    # replace <node-to-drain> with the name of your node you are draining
kubectl drain <node-to-drain> --ignore-daemonsets

    You should see output similar to this:

    node/ip-172-31-85-18 cordoned
node/ip-172-31-85-18 drained

Upgrade the kubelet configuration

  1. From the Windows node, call the following command to sync new kubelet configuration:

    kubeadm upgrade node

Upgrade kubelet and kube-proxy

  1. From the Windows node, upgrade and restart the kubelet:

    stop-service kubelet
curl.exe -Lo <path-to-kubelet.exe> "https://dl.k8s.io/v1.32.0/bin/windows/amd64/kubelet.exe"
restart-service kubelet

  2. From the Windows node, upgrade and restart the kube-proxy.

    stop-service kube-proxy
curl.exe -Lo <path-to-kube-proxy.exe> "https://dl.k8s.io/v1.32.0/bin/windows/amd64/kube-proxy.exe"
restart-service kube-proxy

Uncordon the node

  1. From a machine with access to the Kubernetes API, bring the node back online by marking it schedulable:

    # replace <node-to-drain> with the name of your node
kubectl uncordon <node-to-drain>

What's next

1.6 - Configuring a cgroup driver

This page explains how to configure the kubelet's cgroup driver to match the container runtime cgroup driver for kubeadm clusters.

Before you begin

You should be familiar with the Kubernetes container runtime requirements.

Configuring the container runtime cgroup driver

The Container runtimes page explains that the systemd driver is recommended for kubeadm based setups instead of the kubelet's default cgroupfs driver, because kubeadm manages the kubelet as a systemd service.

The page also provides details on how to set up a number of different container runtimes with the systemd driver by default.

Configuring the kubelet cgroup driver

kubeadm allows you to pass a KubeletConfiguration structure during kubeadm init. This KubeletConfiguration can include the cgroupDriver field which controls the cgroup driver of the kubelet.

A minimal example of configuring the field explicitly:

# kubeadm-config.yaml
kind: ClusterConfiguration
apiVersion: kubeadm.k8s.io/v1beta4
kubernetesVersion: v1.21.0
---
kind: KubeletConfiguration
apiVersion: kubelet.config.k8s.io/v1beta1
cgroupDriver: systemd

Such a configuration file can then be passed to the kubeadm command:

kubeadm init --config kubeadm-config.yaml

Using the cgroupfs driver

To use cgroupfs and to prevent kubeadm upgrade from modifying the KubeletConfiguration cgroup driver on existing setups, you must be explicit about its value. This applies to a case where you do not wish future versions of kubeadm to apply the systemd driver by default.

See the below section on "Modify the kubelet ConfigMap" for details on how to be explicit about the value.

If you wish to configure a container runtime to use the cgroupfs driver, you must refer to the documentation of the container runtime of your choice.

Migrating to the systemd driver

To change the cgroup driver of an existing kubeadm cluster from cgroupfs to systemd in-place, a similar procedure to a kubelet upgrade is required. This must include both steps outlined below.

Modify the kubelet ConfigMap

  • Call kubectl edit cm kubelet-config -n kube-system.

  • Either modify the existing cgroupDriver value or add a new field that looks like this:

    cgroupDriver: systemd

    This field must be present under the kubelet: section of the ConfigMap.

Update the cgroup driver on all nodes

For each node in the cluster:

  • Drain the node using kubectl drain <node-name> --ignore-daemonsets
  • Stop the kubelet using systemctl stop kubelet
  • Stop the container runtime
  • Modify the container runtime cgroup driver to systemd
  • Set cgroupDriver: systemd in /var/lib/kubelet/config.yaml
  • Start the container runtime
  • Start the kubelet using systemctl start kubelet
  • Uncordon the node using kubectl uncordon <node-name>

Execute these steps on nodes one at a time to ensure workloads have sufficient time to schedule on different nodes.

Once the process is complete ensure that all nodes and workloads are healthy.

1.7 - Certificate Management with kubeadm

FEATURE STATE: Kubernetes v1.15 [stable]

Client certificates generated by kubeadm expire after 1 year. This page explains how to manage certificate renewals with kubeadm. It also covers other tasks related to kubeadm certificate management.

The Kubernetes project recommends upgrading to the latest patch releases promptly, and to ensure that you are running a supported minor release of Kubernetes. Following this recommendation helps you to to stay secure.

Before you begin

You should be familiar with PKI certificates and requirements in Kubernetes.

This guide covers the usage of the openssl command (used for manual certificate signing, if you choose that approach), but you can use your preferred tools.

Some of the steps here use sudo for administrator access. You can use any equivalent tool.

Using custom certificates

By default, kubeadm generates all the certificates needed for a cluster to run. You can override this behavior by providing your own certificates.

To do so, you must place them in whatever directory is specified by the --cert-dir flag or the certificatesDir field of kubeadm's ClusterConfiguration. By default this is /etc/kubernetes/pki.

If a given certificate and private key pair exists before running kubeadm init, kubeadm does not overwrite them. This means you can, for example, copy an existing CA into /etc/kubernetes/pki/ca.crt and /etc/kubernetes/pki/ca.key, and kubeadm will use this CA for signing the rest of the certificates.

External CA mode

It is also possible to provide only the ca.crt file and not the ca.key file (this is only available for the root CA file, not other cert pairs). If all other certificates and kubeconfig files are in place, kubeadm recognizes this condition and activates the "External CA" mode. kubeadm will proceed without the CA key on disk.

Instead, run the controller-manager standalone with --controllers=csrsigner and point to the CA certificate and key.

There are various ways to prepare the component credentials when using external CA mode.

Manual preparation of component credentials

PKI certificates and requirements includes information on how to prepare all the required by kubeadm component credentials manually.

This guide covers the usage of the openssl command (used for manual certificate signing, if you choose that approach), but you can use your preferred tools.

Preparation of credentials by signing CSRs generated by kubeadm

kubeadm can generate CSR files that you can sign manually with tools like openssl and your external CA. These CSR files will include all the specification for credentials that components deployed by kubeadm require.

Automated preparation of component credentials by using kubeadm phases

Alternatively, it is possible to use kubeadm phase commands to automate this process.

  • Go to a host that you want to prepare as a kubeadm control plane node with external CA.
  • Copy the external CA files ca.crt and ca.key that you have into /etc/kubernetes/pki on the node.
  • Prepare a temporary kubeadm configuration file called config.yaml that can be used with kubeadm init. Make sure that this file includes any relevant cluster wide or host-specific information that could be included in certificates, such as, ClusterConfiguration.controlPlaneEndpoint, ClusterConfiguration.certSANs and InitConfiguration.APIEndpoint.
  • On the same host execute the commands kubeadm init phase kubeconfig all --config config.yaml and kubeadm init phase certs all --config config.yaml. This will generate all required kubeconfig files and certificates under /etc/kubernetes/ and its pki sub directory.
  • Inspect the generated files. Delete /etc/kubernetes/pki/ca.key, delete or move to a safe location the file /etc/kubernetes/super-admin.conf.
  • On nodes where kubeadm join will be called also delete /etc/kubernetes/kubelet.conf. This file is only required on the first node where kubeadm init will be called.
  • Note that some files such pki/sa.*, pki/front-proxy-ca.* and pki/etc/ca.* are shared between control plane nodes, You can generate them once and distribute them manually to nodes where kubeadm join will be called, or you can use the --upload-certs functionality of kubeadm init and --certificate-key of kubeadm join to automate this distribution.

Once the credentials are prepared on all nodes, call kubeadm init and kubeadm join for these nodes to join the cluster. kubeadm will use the existing kubeconfig and certificate files under /etc/kubernetes/ and its pki sub directory.

Certificate expiry and management

You can use the check-expiration subcommand to check when certificates expire:

kubeadm certs check-expiration

The output is similar to this:

CERTIFICATE                EXPIRES                  RESIDUAL TIME   CERTIFICATE AUTHORITY   EXTERNALLY MANAGED
admin.conf                 Dec 30, 2020 23:36 UTC   364d                                    no
apiserver                  Dec 30, 2020 23:36 UTC   364d            ca                      no
apiserver-etcd-client      Dec 30, 2020 23:36 UTC   364d            etcd-ca                 no
apiserver-kubelet-client   Dec 30, 2020 23:36 UTC   364d            ca                      no
controller-manager.conf    Dec 30, 2020 23:36 UTC   364d                                    no
etcd-healthcheck-client    Dec 30, 2020 23:36 UTC   364d            etcd-ca                 no
etcd-peer                  Dec 30, 2020 23:36 UTC   364d            etcd-ca                 no
etcd-server                Dec 30, 2020 23:36 UTC   364d            etcd-ca                 no
front-proxy-client         Dec 30, 2020 23:36 UTC   364d            front-proxy-ca          no
scheduler.conf             Dec 30, 2020 23:36 UTC   364d                                    no

CERTIFICATE AUTHORITY   EXPIRES                  RESIDUAL TIME   EXTERNALLY MANAGED
ca                      Dec 28, 2029 23:36 UTC   9y              no
etcd-ca                 Dec 28, 2029 23:36 UTC   9y              no
front-proxy-ca          Dec 28, 2029 23:36 UTC   9y              no

The command shows expiration/residual time for the client certificates in the /etc/kubernetes/pki folder and for the client certificate embedded in the kubeconfig files used by kubeadm (admin.conf, controller-manager.conf and scheduler.conf).

Additionally, kubeadm informs the user if the certificate is externally managed; in this case, the user should take care of managing certificate renewal manually/using other tools.

The kubelet.conf configuration file is not included in the list above because kubeadm configures kubelet for automatic certificate renewal with rotatable certificates under /var/lib/kubelet/pki. To repair an expired kubelet client certificate see Kubelet client certificate rotation fails.

Automatic certificate renewal

kubeadm renews all the certificates during control plane upgrade.

This feature is designed for addressing the simplest use cases; if you don't have specific requirements on certificate renewal and perform Kubernetes version upgrades regularly (less than 1 year in between each upgrade), kubeadm will take care of keeping your cluster up to date and reasonably secure.

If you have more complex requirements for certificate renewal, you can opt out from the default behavior by passing --certificate-renewal=false to kubeadm upgrade apply or to kubeadm upgrade node.

Manual certificate renewal

You can renew your certificates manually at any time with the kubeadm certs renew command, with the appropriate command line options. If you are running cluster with a replicated control plane, this command needs to be executed on all the control-plane nodes.

This command performs the renewal using CA (or front-proxy-CA) certificate and key stored in /etc/kubernetes/pki.

kubeadm certs renew uses the existing certificates as the authoritative source for attributes (Common Name, Organization, subject alternative name) and does not rely on the kubeadm-config ConfigMap. Even so, the Kubernetes project recommends keeping the served certificate and the associated values in that ConfigMap synchronized, to avoid any risk of confusion.

After running the command you should restart the control plane Pods. This is required since dynamic certificate reload is currently not supported for all components and certificates. Static Pods are managed by the local kubelet and not by the API Server, thus kubectl cannot be used to delete and restart them. To restart a static Pod you can temporarily remove its manifest file from /etc/kubernetes/manifests/ and wait for 20 seconds (see the fileCheckFrequency value in KubeletConfiguration struct. The kubelet will terminate the Pod if it's no longer in the manifest directory. You can then move the file back and after another fileCheckFrequency period, the kubelet will recreate the Pod and the certificate renewal for the component can complete.

kubeadm certs renew can renew any specific certificate or, with the subcommand all, it can renew all of them:

# If you are running cluster with a replicated control plane, this command
# needs to be executed on all the control-plane nodes.
kubeadm certs renew all

Copying the administrator certificate (optional)

Clusters built with kubeadm often copy the admin.conf certificate into $HOME/.kube/config, as instructed in Creating a cluster with kubeadm. On such a system, to update the contents of $HOME/.kube/config after renewing the admin.conf, you could run the following commands:

sudo cp -i /etc/kubernetes/admin.conf $HOME/.kube/config
sudo chown $(id -u):$(id -g) $HOME/.kube/config

Renew certificates with the Kubernetes certificates API

This section provides more details about how to execute manual certificate renewal using the Kubernetes certificates API.

Set up a signer

The Kubernetes Certificate Authority does not work out of the box. You can configure an external signer such as cert-manager, or you can use the built-in signer.

The built-in signer is part of kube-controller-manager.

To activate the built-in signer, you must pass the --cluster-signing-cert-file and --cluster-signing-key-file flags.

If you're creating a new cluster, you can use a kubeadm configuration file:

apiVersion: kubeadm.k8s.io/v1beta4
kind: ClusterConfiguration
controllerManager:
  extraArgs:
  - name: "cluster-signing-cert-file"
    value: "/etc/kubernetes/pki/ca.crt"
  - name: "cluster-signing-key-file"
    value: "/etc/kubernetes/pki/ca.key"

Create certificate signing requests (CSR)

See Create CertificateSigningRequest for creating CSRs with the Kubernetes API.

Renew certificates with external CA

This section provide more details about how to execute manual certificate renewal using an external CA.

To better integrate with external CAs, kubeadm can also produce certificate signing requests (CSRs). A CSR represents a request to a CA for a signed certificate for a client. In kubeadm terms, any certificate that would normally be signed by an on-disk CA can be produced as a CSR instead. A CA, however, cannot be produced as a CSR.

Renewal by using certificate signing requests (CSR)

Renewal of ceritficates is possible by generating new CSRs and signing them with the external CA. For more details about working with CSRs generated by kubeadm see the section Signing certificate signing requests (CSR) generated by kubeadm.

Certificate authority (CA) rotation

Kubeadm does not support rotation or replacement of CA certificates out of the box.

For more information about manual rotation or replacement of CA, see manual rotation of CA certificates.

Enabling signed kubelet serving certificates

By default the kubelet serving certificate deployed by kubeadm is self-signed. This means a connection from external services like the metrics-server to a kubelet cannot be secured with TLS.

To configure the kubelets in a new kubeadm cluster to obtain properly signed serving certificates you must pass the following minimal configuration to kubeadm init:

apiVersion: kubeadm.k8s.io/v1beta4
kind: ClusterConfiguration
---
apiVersion: kubelet.config.k8s.io/v1beta1
kind: KubeletConfiguration
serverTLSBootstrap: true

If you have already created the cluster you must adapt it by doing the following:

  • Find and edit the kubelet-config-1.32 ConfigMap in the kube-system namespace. In that ConfigMap, the kubelet key has a KubeletConfiguration document as its value. Edit the KubeletConfiguration document to set serverTLSBootstrap: true.
  • On each node, add the serverTLSBootstrap: true field in /var/lib/kubelet/config.yaml and restart the kubelet with systemctl restart kubelet

The field serverTLSBootstrap: true will enable the bootstrap of kubelet serving certificates by requesting them from the certificates.k8s.io API. One known limitation is that the CSRs (Certificate Signing Requests) for these certificates cannot be automatically approved by the default signer in the kube-controller-manager - kubernetes.io/kubelet-serving. This will require action from the user or a third party controller.

These CSRs can be viewed using:

kubectl get csr

NAME        AGE     SIGNERNAME                        REQUESTOR                      CONDITION
csr-9wvgt   112s    kubernetes.io/kubelet-serving     system:node:worker-1           Pending
csr-lz97v   1m58s   kubernetes.io/kubelet-serving     system:node:control-plane-1    Pending

To approve them you can do the following:

kubectl certificate approve <CSR-name>

By default, these serving certificate will expire after one year. Kubeadm sets the KubeletConfiguration field rotateCertificates to true, which means that close to expiration a new set of CSRs for the serving certificates will be created and must be approved to complete the rotation. To understand more see Certificate Rotation.

If you are looking for a solution for automatic approval of these CSRs it is recommended that you contact your cloud provider and ask if they have a CSR signer that verifies the node identity with an out of band mechanism.

Third party custom controllers can be used:

Such a controller is not a secure mechanism unless it not only verifies the CommonName in the CSR but also verifies the requested IPs and domain names. This would prevent a malicious actor that has access to a kubelet client certificate to create CSRs requesting serving certificates for any IP or domain name.

Generating kubeconfig files for additional users

During cluster creation, kubeadm init signs the certificate in the super-admin.conf to have Subject: O = system:masters, CN = kubernetes-super-admin. system:masters is a break-glass, super user group that bypasses the authorization layer (for example, RBAC). The file admin.conf is also created by kubeadm on control plane nodes and it contains a certificate with Subject: O = kubeadm:cluster-admins, CN = kubernetes-admin. kubeadm:cluster-admins is a group logically belonging to kubeadm. If your cluster uses RBAC (the kubeadm default), the kubeadm:cluster-admins group is bound to the cluster-admin ClusterRole.

You can use the kubeadm kubeconfig user command to generate kubeconfig files for additional users. The command accepts a mixture of command line flags and kubeadm configuration options. The generated kubeconfig will be written to stdout and can be piped to a file using kubeadm kubeconfig user ... > somefile.conf.

Example configuration file that can be used with --config:

# example.yaml
apiVersion: kubeadm.k8s.io/v1beta4
kind: ClusterConfiguration
# Will be used as the target "cluster" in the kubeconfig
clusterName: "kubernetes"
# Will be used as the "server" (IP or DNS name) of this cluster in the kubeconfig
controlPlaneEndpoint: "some-dns-address:6443"
# The cluster CA key and certificate will be loaded from this local directory
certificatesDir: "/etc/kubernetes/pki"

Make sure that these settings match the desired target cluster settings. To see the settings of an existing cluster use:

kubectl get cm kubeadm-config -n kube-system -o=jsonpath="{.data.ClusterConfiguration}"

The following example will generate a kubeconfig file with credentials valid for 24 hours for a new user johndoe that is part of the appdevs group:

kubeadm kubeconfig user --config example.yaml --org appdevs --client-name johndoe --validity-period 24h

The following example will generate a kubeconfig file with administrator credentials valid for 1 week:

kubeadm kubeconfig user --config example.yaml --client-name admin --validity-period 168h

Signing certificate signing requests (CSR) generated by kubeadm

You can create certificate signing requests with kubeadm certs generate-csr. Calling this command will generate .csr / .key file pairs for regular certificates. For certificates embedded in kubeconfig files, the command will generate a .csr / .conf pair where the key is already embedded in the .conf file.

A CSR file contains all relevant information for a CA to sign a certificate. kubeadm uses a well defined specification for all its certificates and CSRs.

The default certificate directory is /etc/kubernetes/pki, while the default directory for kubeconfig files is /etc/kubernetes. These defaults can be overridden with the flags --cert-dir and --kubeconfig-dir, respectively.

To pass custom options to kubeadm certs generate-csr use the --config flag, which accepts a kubeadm configuration file, similarly to commands such as kubeadm init. Any specification such as extra SANs and custom IP addresses must be stored in the same configuration file and used for all relevant kubeadm commands by passing it as --config.

Preparing CA and service account files

On the primary control plane node, where kubeadm init will be executed, call the following commands:

sudo kubeadm init phase certs ca
sudo kubeadm init phase certs etcd-ca
sudo kubeadm init phase certs front-proxy-ca
sudo kubeadm init phase certs sa

This will populate the folders /etc/kubernetes/pki and /etc/kubernetes/pki/etcd with all self-signed CA files (certificates and keys) and service account (public and private keys) that kubeadm needs for a control plane node.

For secondary control plane nodes (kubeadm join --control-plane) there is no need to call the above commands. Depending on how you setup the High Availability cluster, you either have to manually copy the same files from the primary control plane node, or use the automated --upload-certs functionality of kubeadm init.

Generate CSRs

The kubeadm certs generate-csr command generates CSRs for all known certificates managed by kubeadm. Once the command is done you must manually delete .csr, .conf or .key files that you don't need.

Considerations for kubelet.conf

This section applies to both control plane and worker nodes.

If you have deleted the ca.key file from control plane nodes (External CA mode), the active kube-controller-manager in this cluster will not be able to sign kubelet client certificates. If no external method for signing these certificates exists in your setup (such as an external signer, you could manually sign the kubelet.conf.csr as explained in this guide.

Note that this also means that the automatic kubelet client certificate rotation will be disabled. If so, close to certificate expiration, you must generate a new kubelet.conf.csr, sign the certificate, embed it in kubelet.conf and restart the kubelet.

If this does not apply to your setup, you can skip processing the kubelet.conf.csr on secondary control plane and on workers nodes (all nodes that call kubeadm join ...). That is because the active kube-controller-manager will be responsible for signing new kubelet client certificates.

Control plane nodes

Execute the following command on primary (kubeadm init) and secondary (kubeadm join --control-plane) control plane nodes to generate all CSR files:

sudo kubeadm certs generate-csr

If external etcd is to be used, follow the External etcd with kubeadm guide to understand what CSR files are needed on the kubeadm and etcd nodes. Other .csr and .key files under /etc/kubernetes/pki/etcd can be removed.

Based on the explanation in Considerations for kubelet.conf keep or delete the kubelet.conf and kubelet.conf.csr files.

Worker nodes

Based on the explanation in Considerations for kubelet.conf, optionally call:

sudo kubeadm certs generate-csr

and keep only the kubelet.conf and kubelet.conf.csr files. Alternatively skip the steps for worker nodes entirely.

Signing CSRs for all certificates

Repeat this step for all nodes that have CSR files.

Write the following script in the /etc/kubernetes directory, navigate to the directory and execute the script. The script will generate certificates for all CSR files that are present in the /etc/kubernetes tree.

#!/bin/bash

# Set certificate expiration time in days
DAYS=365

# Process all CSR files except those for front-proxy and etcd
find ./ -name "*.csr" | grep -v "pki/etcd" | grep -v "front-proxy" | while read -r FILE;
do
    echo "* Processing ${FILE} ..."
    FILE=${FILE%.*} # Trim the extension
    if [ -f "./pki/ca.srl" ]; then
        SERIAL_FLAG="-CAserial ./pki/ca.srl"
    else
        SERIAL_FLAG="-CAcreateserial"
    fi
    openssl x509 -req -days "${DAYS}" -CA ./pki/ca.crt -CAkey ./pki/ca.key ${SERIAL_FLAG} \
        -in "${FILE}.csr" -out "${FILE}.crt"
    sleep 2
done

# Process all etcd CSRs
find ./pki/etcd -name "*.csr" | while read -r FILE;
do
    echo "* Processing ${FILE} ..."
    FILE=${FILE%.*} # Trim the extension
    if [ -f "./pki/etcd/ca.srl" ]; then
        SERIAL_FLAG=-CAserial ./pki/etcd/ca.srl
    else
        SERIAL_FLAG=-CAcreateserial
    fi
    openssl x509 -req -days "${DAYS}" -CA ./pki/etcd/ca.crt -CAkey ./pki/etcd/ca.key ${SERIAL_FLAG} \
        -in "${FILE}.csr" -out "${FILE}.crt"
done

# Process front-proxy CSRs
echo "* Processing ./pki/front-proxy-client.csr ..."
openssl x509 -req -days "${DAYS}" -CA ./pki/front-proxy-ca.crt -CAkey ./pki/front-proxy-ca.key -CAcreateserial \
    -in ./pki/front-proxy-client.csr -out ./pki/front-proxy-client.crt

Embedding certificates in kubeconfig files

Repeat this step for all nodes that have CSR files.

Write the following script in the /etc/kubernetes directory, navigate to the directory and execute the script. The script will take the .crt files that were signed for kubeconfig files from CSRs in the previous step and will embed them in the kubeconfig files.

#!/bin/bash

CLUSTER=kubernetes
find ./ -name "*.conf" | while read -r FILE;
do
    echo "* Processing ${FILE} ..."
    KUBECONFIG="${FILE}" kubectl config set-cluster "${CLUSTER}" --certificate-authority ./pki/ca.crt --embed-certs
    USER=$(KUBECONFIG="${FILE}" kubectl config view -o jsonpath='{.users[0].name}')
    KUBECONFIG="${FILE}" kubectl config set-credentials "${USER}" --client-certificate "${FILE}.crt" --embed-certs
done

Performing cleanup

Perform this step on all nodes that have CSR files.

Write the following script in the /etc/kubernetes directory, navigate to the directory and execute the script.

#!/bin/bash

# Cleanup CSR files
rm -f ./*.csr ./pki/*.csr ./pki/etcd/*.csr # Clean all CSR files

# Cleanup CRT files that were already embedded in kubeconfig files
rm -f ./*.crt

Optionally, move .srl files to the next node to be processed.

Optionally, if using external CA remove the /etc/kubernetes/pki/ca.key file, as explained in the External CA node section.

kubeadm node initialization

Once CSR files have been signed and required certificates are in place on the hosts you want to use as nodes, you can use the commands kubeadm init and kubeadm join to create a Kubernetes cluster from these nodes. During init and join, kubeadm uses existing certificates, encryption keys and kubeconfig files that it finds in the /etc/kubernetes tree on the host's local filesystem.

1.8 - Reconfiguring a kubeadm cluster

kubeadm does not support automated ways of reconfiguring components that were deployed on managed nodes. One way of automating this would be by using a custom operator.

To modify the components configuration you must manually edit associated cluster objects and files on disk.

This guide shows the correct sequence of steps that need to be performed to achieve kubeadm cluster reconfiguration.

Before you begin

  • You need a cluster that was deployed using kubeadm
  • Have administrator credentials (/etc/kubernetes/admin.conf) and network connectivity to a running kube-apiserver in the cluster from a host that has kubectl installed
  • Have a text editor installed on all hosts

Reconfiguring the cluster

kubeadm writes a set of cluster wide component configuration options in ConfigMaps and other objects. These objects must be manually edited. The command kubectl edit can be used for that.

The kubectl edit command will open a text editor where you can edit and save the object directly.

You can use the environment variables KUBECONFIG and KUBE_EDITOR to specify the location of the kubectl consumed kubeconfig file and preferred text editor.

For example:

KUBECONFIG=/etc/kubernetes/admin.conf KUBE_EDITOR=nano kubectl edit <parameters>

Applying cluster configuration changes

Updating the ClusterConfiguration

During cluster creation and upgrade, kubeadm writes its ClusterConfiguration in a ConfigMap called kubeadm-config in the kube-system namespace.

To change a particular option in the ClusterConfiguration you can edit the ConfigMap with this command:

kubectl edit cm -n kube-system kubeadm-config

The configuration is located under the data.ClusterConfiguration key.

Reflecting ClusterConfiguration changes on control plane nodes

kubeadm manages the control plane components as static Pod manifests located in the directory /etc/kubernetes/manifests. Any changes to the ClusterConfiguration under the apiServer, controllerManager, scheduler or etcd keys must be reflected in the associated files in the manifests directory on a control plane node.

Such changes may include:

  • extraArgs - requires updating the list of flags passed to a component container
  • extraVolumes - requires updating the volume mounts for a component container
  • *SANs - requires writing new certificates with updated Subject Alternative Names

Before proceeding with these changes, make sure you have backed up the directory /etc/kubernetes/.

To write new certificates you can use:

kubeadm init phase certs <component-name> --config <config-file>

To write new manifest files in /etc/kubernetes/manifests you can use:

# For Kubernetes control plane components
kubeadm init phase control-plane <component-name> --config <config-file>
# For local etcd
kubeadm init phase etcd local --config <config-file>

The <config-file> contents must match the updated ClusterConfiguration. The <component-name> value must be a name of a Kubernetes control plane component (apiserver, controller-manager or scheduler).

Applying kubelet configuration changes

Updating the KubeletConfiguration

During cluster creation and upgrade, kubeadm writes its KubeletConfiguration in a ConfigMap called kubelet-config in the kube-system namespace.

You can edit the ConfigMap with this command:

kubectl edit cm -n kube-system kubelet-config

The configuration is located under the data.kubelet key.

Reflecting the kubelet changes

To reflect the change on kubeadm nodes you must do the following:

  • Log in to a kubeadm node
  • Run kubeadm upgrade node phase kubelet-config to download the latest kubelet-config ConfigMap contents into the local file /var/lib/kubelet/config.yaml
  • Edit the file /var/lib/kubelet/kubeadm-flags.env to apply additional configuration with flags
  • Restart the kubelet service with systemctl restart kubelet

Applying kube-proxy configuration changes

Updating the KubeProxyConfiguration

During cluster creation and upgrade, kubeadm writes its KubeProxyConfiguration in a ConfigMap in the kube-system namespace called kube-proxy.

This ConfigMap is used by the kube-proxy DaemonSet in the kube-system namespace.

To change a particular option in the KubeProxyConfiguration, you can edit the ConfigMap with this command:

kubectl edit cm -n kube-system kube-proxy

The configuration is located under the data.config.conf key.

Reflecting the kube-proxy changes

Once the kube-proxy ConfigMap is updated, you can restart all kube-proxy Pods:

Obtain the Pod names:

kubectl get po -n kube-system | grep kube-proxy

Delete a Pod with:

kubectl delete po -n kube-system <pod-name>

New Pods that use the updated ConfigMap will be created.

Applying CoreDNS configuration changes

Updating the CoreDNS Deployment and Service

kubeadm deploys CoreDNS as a Deployment called coredns and with a Service kube-dns, both in the kube-system namespace.

To update any of the CoreDNS settings, you can edit the Deployment and Service objects:

kubectl edit deployment -n kube-system coredns
kubectl edit service -n kube-system kube-dns

Reflecting the CoreDNS changes

Once the CoreDNS changes are applied you can delete the CoreDNS Pods:

Obtain the Pod names:

kubectl get po -n kube-system | grep coredns

Delete a Pod with:

kubectl delete po -n kube-system <pod-name>

New Pods with the updated CoreDNS configuration will be created.

Persisting the reconfiguration

During the execution of kubeadm upgrade on a managed node, kubeadm might overwrite configuration that was applied after the cluster was created (reconfiguration).

Persisting Node object reconfiguration

kubeadm writes Labels, Taints, CRI socket and other information on the Node object for a particular Kubernetes node. To change any of the contents of this Node object you can use:

kubectl edit no <node-name>

During kubeadm upgrade the contents of such a Node might get overwritten. If you would like to persist your modifications to the Node object after upgrade, you can prepare a kubectl patch and apply it to the Node object:

kubectl patch no <node-name> --patch-file <patch-file>

Persisting control plane component reconfiguration

The main source of control plane configuration is the ClusterConfiguration object stored in the cluster. To extend the static Pod manifests configuration, patches can be used.

These patch files must remain as files on the control plane nodes to ensure that they can be used by the kubeadm upgrade ... --patches <directory>.

If reconfiguration is done to the ClusterConfiguration and static Pod manifests on disk, the set of node specific patches must be updated accordingly.

Persisting kubelet reconfiguration

Any changes to the KubeletConfiguration stored in /var/lib/kubelet/config.yaml will be overwritten on kubeadm upgrade by downloading the contents of the cluster wide kubelet-config ConfigMap. To persist kubelet node specific configuration either the file /var/lib/kubelet/config.yaml has to be updated manually post-upgrade or the file /var/lib/kubelet/kubeadm-flags.env can include flags. The kubelet flags override the associated KubeletConfiguration options, but note that some of the flags are deprecated.

A kubelet restart will be required after changing /var/lib/kubelet/config.yaml or /var/lib/kubelet/kubeadm-flags.env.

What's next

1.9 - Changing The Kubernetes Package Repository

This page explains how to enable a package repository for the desired Kubernetes minor release upon upgrading a cluster. This is only needed for users of the community-owned package repositories hosted at pkgs.k8s.io. Unlike the legacy package repositories, the community-owned package repositories are structured in a way that there's a dedicated package repository for each Kubernetes minor version.

Before you begin

This document assumes that you're already using the community-owned package repositories (pkgs.k8s.io). If that's not the case, it's strongly recommended to migrate to the community-owned package repositories as described in the official announcement.

Verifying if the Kubernetes package repositories are used

If you're unsure whether you're using the community-owned package repositories or the legacy package repositories, take the following steps to verify:

Print the contents of the file that defines the Kubernetes apt repository:

# On your system, this configuration file could have a different name
pager /etc/apt/sources.list.d/kubernetes.list

If you see a line similar to:

deb [signed-by=/etc/apt/keyrings/kubernetes-apt-keyring.gpg] https://pkgs.k8s.io/core:/stable:/v1.31/deb/ /

You're using the Kubernetes package repositories and this guide applies to you. Otherwise, it's strongly recommended to migrate to the Kubernetes package repositories as described in the official announcement.

Print the contents of the file that defines the Kubernetes yum repository:

# On your system, this configuration file could have a different name
cat /etc/yum.repos.d/kubernetes.repo

If you see a baseurl similar to the baseurl in the output below:

[kubernetes]
name=Kubernetes
baseurl=https://pkgs.k8s.io/core:/stable:/v1.31/rpm/
enabled=1
gpgcheck=1
gpgkey=https://pkgs.k8s.io/core:/stable:/v1.31/rpm/repodata/repomd.xml.key
exclude=kubelet kubeadm kubectl

You're using the Kubernetes package repositories and this guide applies to you. Otherwise, it's strongly recommended to migrate to the Kubernetes package repositories as described in the official announcement.

Print the contents of the file that defines the Kubernetes zypper repository:

# On your system, this configuration file could have a different name
cat /etc/zypp/repos.d/kubernetes.repo

If you see a baseurl similar to the baseurl in the output below:

[kubernetes]
name=Kubernetes
baseurl=https://pkgs.k8s.io/core:/stable:/v1.31/rpm/
enabled=1
gpgcheck=1
gpgkey=https://pkgs.k8s.io/core:/stable:/v1.31/rpm/repodata/repomd.xml.key
exclude=kubelet kubeadm kubectl

You're using the Kubernetes package repositories and this guide applies to you. Otherwise, it's strongly recommended to migrate to the Kubernetes package repositories as described in the official announcement.

Switching to another Kubernetes package repository

This step should be done upon upgrading from one to another Kubernetes minor release in order to get access to the packages of the desired Kubernetes minor version.

  1. Open the file that defines the Kubernetes apt repository using a text editor of your choice:

    nano /etc/apt/sources.list.d/kubernetes.list

    You should see a single line with the URL that contains your current Kubernetes minor version. For example, if you're using v1.31, you should see this:

    deb [signed-by=/etc/apt/keyrings/kubernetes-apt-keyring.gpg] https://pkgs.k8s.io/core:/stable:/v1.31/deb/ /

  2. Change the version in the URL to the next available minor release, for example:

    deb [signed-by=/etc/apt/keyrings/kubernetes-apt-keyring.gpg] https://pkgs.k8s.io/core:/stable:/v1.32/deb/ /

  3. Save the file and exit your text editor. Continue following the relevant upgrade instructions.

  1. Open the file that defines the Kubernetes yum repository using a text editor of your choice:

    nano /etc/yum.repos.d/kubernetes.repo

    You should see a file with two URLs that contain your current Kubernetes minor version. For example, if you're using v1.31, you should see this:

    [kubernetes]
name=Kubernetes
baseurl=https://pkgs.k8s.io/core:/stable:/v1.31/rpm/
enabled=1
gpgcheck=1
gpgkey=https://pkgs.k8s.io/core:/stable:/v1.31/rpm/repodata/repomd.xml.key
exclude=kubelet kubeadm kubectl cri-tools kubernetes-cni

  2. Change the version in these URLs to the next available minor release, for example:

    [kubernetes]
name=Kubernetes
baseurl=https://pkgs.k8s.io/core:/stable:/v1.32/rpm/
enabled=1
gpgcheck=1
gpgkey=https://pkgs.k8s.io/core:/stable:/v1.32/rpm/repodata/repomd.xml.key
exclude=kubelet kubeadm kubectl cri-tools kubernetes-cni

  3. Save the file and exit your text editor. Continue following the relevant upgrade instructions.

What's next

2 - Overprovision Node Capacity For A Cluster

This page guides you through configuring Node overprovisioning in your Kubernetes cluster. Node overprovisioning is a strategy that proactively reserves a portion of your cluster's compute resources. This reservation helps reduce the time required to schedule new pods during scaling events, enhancing your cluster's responsiveness to sudden spikes in traffic or workload demands.

By maintaining some unused capacity, you ensure that resources are immediately available when new pods are created, preventing them from entering a pending state while the cluster scales up.

Before you begin

  • You need to have a Kubernetes cluster, and the kubectl command-line tool must be configured to communicate with your cluster.
  • You should already have a basic understanding of Deployments, Pod priority, and PriorityClasses.
  • Your cluster must be set up with an autoscaler that manages nodes based on demand.

Create a PriorityClass

Begin by defining a PriorityClass for the placeholder Pods. First, create a PriorityClass with a negative priority value, that you will shortly assign to the placeholder pods. Later, you will set up a Deployment that uses this PriorityClass

priorityclass/low-priority-class.yaml 
apiVersion: scheduling.k8s.io/v1
kind: PriorityClass
metadata:
  name: placeholder # these Pods represent placeholder capacity
value: -1000
globalDefault: false
description: "Negative priority for placeholder pods to enable overprovisioning."

Then create the PriorityClass:

kubectl apply -f https://k8s.io/examples/priorityclass/low-priority-class.yaml

You will next define a Deployment that uses the negative-priority PriorityClass and runs a minimal container. When you add this to your cluster, Kubernetes runs those placeholder pods to reserve capacity. Any time there is a capacity shortage, the control plane will pick one these placeholder pods as the first candidate to preempt.

Run Pods that request node capacity

Review the sample manifest:

deployments/deployment-with-capacity-reservation.yaml 
apiVersion: apps/v1
kind: Deployment
metadata:
  name: capacity-reservation
  # You should decide what namespace to deploy this into
spec:
  replicas: 1
  selector:
    matchLabels:
      app.kubernetes.io/name: capacity-placeholder
  template:
    metadata:
      labels:
        app.kubernetes.io/name: capacity-placeholder
      annotations:
        kubernetes.io/description: "Capacity reservation"
    spec:
      priorityClassName: placeholder
      affinity: # Try to place these overhead Pods on different nodes
                # if possible
        podAntiAffinity:
          preferredDuringSchedulingIgnoredDuringExecution:
          - labelSelector:
              matchLabels:
                app: placeholder
            topologyKey: "kubernetes.io/hostname"
      containers:
      - name: pause
        image: registry.k8s.io/pause:3.6
        resources:
          requests:
            cpu: "50m"
            memory: "512Mi"
          limits:
            memory: "512Mi"

Pick a namespace for the placeholder pods

You should select, or create, a namespace that the placeholder Pods will go into.

Create the placeholder deployment

Create a Deployment based on that manifest:

# Change the namespace name "example"
kubectl --namespace example apply -f https://k8s.io/examples/deployments/deployment-with-capacity-reservation.yaml

Adjust placeholder resource requests

Configure the resource requests and limits for the placeholder pods to define the amount of overprovisioned resources you want to maintain. This reservation ensures that a specific amount of CPU and memory is kept available for new pods.

To edit the Deployment, modify the resources section in the Deployment manifest file to set appropriate requests and limits. You can download that file locally and then edit it with whichever text editor you prefer.

You can also edit the Deployment using kubectl:

kubectl edit deployment capacity-reservation

For example, to reserve a total of a 0.5 CPU and 1GiB of memory across 5 placeholder pods, define the resource requests and limits for a single placeholder pod as follows:

  resources:
    requests:
      cpu: "100m"
      memory: "200Mi"
    limits:
      cpu: "100m"

Set the desired replica count

Calculate the total reserved resources

For example, with 5 replicas each reserving 0.1 CPU and 200MiB of memory:
Total CPU reserved: 5 × 0.1 = 0.5 (in the Pod specification, you'll write the quantity 500m)
Total memory reserved: 5 × 200MiB = 1GiB (in the Pod specification, you'll write 1 Gi)

To scale the Deployment, adjust the number of replicas based on your cluster's size and expected workload:

kubectl scale deployment capacity-reservation --replicas=5

Verify the scaling:

kubectl get deployment capacity-reservation

The output should reflect the updated number of replicas:

NAME                   READY   UP-TO-DATE   AVAILABLE   AGE
capacity-reservation   5/5     5            5           2m

What's next

  • Learn more about PriorityClasses and how they affect pod scheduling.
  • Explore node autoscaling to dynamically adjust your cluster's size based on workload demands.
  • Understand Pod preemption, a key mechanism for Kubernetes to handle resource contention. The same page covers eviction, which is less relevant to the placeholder Pod approach, but is also a mechanism for Kubernetes to react when resources are contended.

3 - Migrating from dockershim

This section presents information you need to know when migrating from dockershim to other container runtimes.

Since the announcement of dockershim deprecation in Kubernetes 1.20, there were questions on how this will affect various workloads and Kubernetes installations. Our Dockershim Removal FAQ is there to help you to understand the problem better.

Dockershim was removed from Kubernetes with the release of v1.24. If you use Docker Engine via dockershim as your container runtime and wish to upgrade to v1.24, it is recommended that you either migrate to another runtime or find an alternative means to obtain Docker Engine support. Check out the container runtimes section to know your options.

The version of Kubernetes with dockershim (1.23) is out of support and the v1.24 will run out of support soon. Make sure to report issues you encountered with the migration so the issues can be fixed in a timely manner and your cluster would be ready for dockershim removal. After v1.24 running out of support, you will need to contact your Kubernetes provider for support or upgrade multiple versions at a time if there are critical issues affecting your cluster.

Your cluster might have more than one kind of node, although this is not a common configuration.

These tasks will help you to migrate:

What's next

  • Check out container runtimes to understand your options for an alternative.
  • If you find a defect or other technical concern relating to migrating away from dockershim, you can report an issue to the Kubernetes project.

3.1 - Changing the Container Runtime on a Node from Docker Engine to containerd

This task outlines the steps needed to update your container runtime to containerd from Docker. It is applicable for cluster operators running Kubernetes 1.23 or earlier. This also covers an example scenario for migrating from dockershim to containerd. Alternative container runtimes can be picked from this page.

Before you begin

Install containerd. For more information see containerd's installation documentation and for specific prerequisite follow the containerd guide.

Drain the node

kubectl drain <node-to-drain> --ignore-daemonsets

Replace <node-to-drain> with the name of your node you are draining.

Stop the Docker daemon

systemctl stop kubelet
systemctl disable docker.service --now

Install Containerd

Follow the guide for detailed steps to install containerd.

  1. Install the containerd.io package from the official Docker repositories. Instructions for setting up the Docker repository for your respective Linux distribution and installing the containerd.io package can be found at Getting started with containerd.

  2. Configure containerd:

    sudo mkdir -p /etc/containerd
containerd config default | sudo tee /etc/containerd/config.toml

  3. Restart containerd:

    sudo systemctl restart containerd

Start a Powershell session, set $Version to the desired version (ex: $Version="1.4.3"), and then run the following commands:

  1. Download containerd:

    curl.exe -L https://github.com/containerd/containerd/releases/download/v$Version/containerd-$Version-windows-amd64.tar.gz -o containerd-windows-amd64.tar.gz
tar.exe xvf .\containerd-windows-amd64.tar.gz

  2. Extract and configure:

    Copy-Item -Path ".\bin\" -Destination "$Env:ProgramFiles\containerd" -Recurse -Force
cd $Env:ProgramFiles\containerd\
.\containerd.exe config default | Out-File config.toml -Encoding ascii

# Review the configuration. Depending on setup you may want to adjust:
# - the sandbox_image (Kubernetes pause image)
# - cni bin_dir and conf_dir locations
Get-Content config.toml

# (Optional - but highly recommended) Exclude containerd from Windows Defender Scans
Add-MpPreference -ExclusionProcess "$Env:ProgramFiles\containerd\containerd.exe"

  3. Start containerd:

    .\containerd.exe --register-service
Start-Service containerd

Configure the kubelet to use containerd as its container runtime

Edit the file /var/lib/kubelet/kubeadm-flags.env and add the containerd runtime to the flags; --container-runtime-endpoint=unix:///run/containerd/containerd.sock.

Users using kubeadm should be aware that the kubeadm tool stores the CRI socket for each host as an annotation in the Node object for that host. To change it you can execute the following command on a machine that has the kubeadm /etc/kubernetes/admin.conf file.

kubectl edit no <node-name>

This will start a text editor where you can edit the Node object. To choose a text editor you can set the KUBE_EDITOR environment variable.

  • Change the value of kubeadm.alpha.kubernetes.io/cri-socket from /var/run/dockershim.sock to the CRI socket path of your choice (for example unix:///run/containerd/containerd.sock).

    Note that new CRI socket paths must be prefixed with unix:// ideally.

  • Save the changes in the text editor, which will update the Node object.

Restart the kubelet

systemctl start kubelet

Verify that the node is healthy

Run kubectl get nodes -o wide and containerd appears as the runtime for the node we just changed.

Remove Docker Engine

If the node appears healthy, remove Docker.

sudo yum remove docker-ce docker-ce-cli


sudo apt-get purge docker-ce docker-ce-cli


sudo dnf remove docker-ce docker-ce-cli


sudo apt-get purge docker-ce docker-ce-cli

The preceding commands don't remove images, containers, volumes, or customized configuration files on your host. To delete them, follow Docker's instructions to Uninstall Docker Engine.

Uncordon the node

kubectl uncordon <node-to-uncordon>

Replace <node-to-uncordon> with the name of your node you previously drained.

3.2 - Migrate Docker Engine nodes from dockershim to cri-dockerd

This page shows you how to migrate your Docker Engine nodes to use cri-dockerd instead of dockershim. You should follow these steps in these scenarios:

  • You want to switch away from dockershim and still use Docker Engine to run containers in Kubernetes.
  • You want to upgrade to Kubernetes v1.32 and your existing cluster relies on dockershim, in which case you must migrate from dockershim and cri-dockerd is one of your options.

To learn more about the removal of dockershim, read the FAQ page.

What is cri-dockerd?

In Kubernetes 1.23 and earlier, you could use Docker Engine with Kubernetes, relying on a built-in component of Kubernetes named dockershim. The dockershim component was removed in the Kubernetes 1.24 release; however, a third-party replacement, cri-dockerd, is available. The cri-dockerd adapter lets you use Docker Engine through the Container Runtime Interface.

If you want to migrate to cri-dockerd so that you can continue using Docker Engine as your container runtime, you should do the following for each affected node:

  1. Install cri-dockerd.
  2. Cordon and drain the node.
  3. Configure the kubelet to use cri-dockerd.
  4. Restart the kubelet.
  5. Verify that the node is healthy.

Test the migration on non-critical nodes first.

You should perform the following steps for each node that you want to migrate to cri-dockerd.

Before you begin

Cordon and drain the node

  1. Cordon the node to stop new Pods scheduling on it:

    kubectl cordon <NODE_NAME>

    Replace <NODE_NAME> with the name of the node.

  2. Drain the node to safely evict running Pods:

    kubectl drain <NODE_NAME> \
    --ignore-daemonsets

Configure the kubelet to use cri-dockerd

The following steps apply to clusters set up using the kubeadm tool. If you use a different tool, you should modify the kubelet using the configuration instructions for that tool.

  1. Open /var/lib/kubelet/kubeadm-flags.env on each affected node.
  2. Modify the --container-runtime-endpoint flag to unix:///var/run/cri-dockerd.sock.
  3. Modify the --container-runtime flag to remote (unavailable in Kubernetes v1.27 and later).

The kubeadm tool stores the node's socket as an annotation on the Node object in the control plane. To modify this socket for each affected node:

  1. Edit the YAML representation of the Node object:

    KUBECONFIG=/path/to/admin.conf kubectl edit no <NODE_NAME>

    Replace the following:

    • /path/to/admin.conf: the path to the kubectl configuration file, admin.conf.
    • <NODE_NAME>: the name of the node you want to modify.

  2. Change kubeadm.alpha.kubernetes.io/cri-socket from /var/run/dockershim.sock to unix:///var/run/cri-dockerd.sock.

  3. Save the changes. The Node object is updated on save.

Restart the kubelet

systemctl restart kubelet

Verify that the node is healthy

To check whether the node uses the cri-dockerd endpoint, follow the instructions in Find out which runtime you use. The --container-runtime-endpoint flag for the kubelet should be unix:///var/run/cri-dockerd.sock.

Uncordon the node

Uncordon the node to let Pods schedule on it:

kubectl uncordon <NODE_NAME>

What's next

3.3 - Find Out What Container Runtime is Used on a Node

This page outlines steps to find out what container runtime the nodes in your cluster use.

Depending on the way you run your cluster, the container runtime for the nodes may have been pre-configured or you need to configure it. If you're using a managed Kubernetes service, there might be vendor-specific ways to check what container runtime is configured for the nodes. The method described on this page should work whenever the execution of kubectl is allowed.

Before you begin

Install and configure kubectl. See Install Tools section for details.

Find out the container runtime used on a Node

Use kubectl to fetch and show node information:

kubectl get nodes -o wide

The output is similar to the following. The column CONTAINER-RUNTIME outputs the runtime and its version.

For Docker Engine, the output is similar to this:

NAME         STATUS   VERSION    CONTAINER-RUNTIME
node-1       Ready    v1.16.15   docker://19.3.1
node-2       Ready    v1.16.15   docker://19.3.1
node-3       Ready    v1.16.15   docker://19.3.1

If your runtime shows as Docker Engine, you still might not be affected by the removal of dockershim in Kubernetes v1.24. Check the runtime endpoint to see if you use dockershim. If you don't use dockershim, you aren't affected.

For containerd, the output is similar to this:

NAME         STATUS   VERSION   CONTAINER-RUNTIME
node-1       Ready    v1.19.6   containerd://1.4.1
node-2       Ready    v1.19.6   containerd://1.4.1
node-3       Ready    v1.19.6   containerd://1.4.1

Find out more information about container runtimes on Container Runtimes page.

Find out what container runtime endpoint you use

The container runtime talks to the kubelet over a Unix socket using the CRI protocol, which is based on the gRPC framework. The kubelet acts as a client, and the runtime acts as the server. In some cases, you might find it useful to know which socket your nodes use. For example, with the removal of dockershim in Kubernetes v1.24 and later, you might want to know whether you use Docker Engine with dockershim.

You can check which socket you use by checking the kubelet configuration on your nodes.

  1. Read the starting commands for the kubelet process:

    tr \\0 ' ' < /proc/"$(pgrep kubelet)"/cmdline

    If you don't have tr or pgrep, check the command line for the kubelet process manually.

  2. In the output, look for the --container-runtime flag and the --container-runtime-endpoint flag.

    • If your nodes use Kubernetes v1.23 and earlier and these flags aren't present or if the --container-runtime flag is not remote, you use the dockershim socket with Docker Engine. The --container-runtime command line argument is not available in Kubernetes v1.27 and later.
    • If the --container-runtime-endpoint flag is present, check the socket name to find out which runtime you use. For example, unix:///run/containerd/containerd.sock is the containerd endpoint.

If you want to change the Container Runtime on a Node from Docker Engine to containerd, you can find out more information on migrating from Docker Engine to containerd, or, if you want to continue using Docker Engine in Kubernetes v1.24 and later, migrate to a CRI-compatible adapter like cri-dockerd.

3.4 - Troubleshooting CNI plugin-related errors

To avoid CNI plugin-related errors, verify that you are using or upgrading to a container runtime that has been tested to work correctly with your version of Kubernetes.

About the "Incompatible CNI versions" and "Failed to destroy network for sandbox" errors

Service issues exist for pod CNI network setup and tear down in containerd v1.6.0-v1.6.3 when the CNI plugins have not been upgraded and/or the CNI config version is not declared in the CNI config files. The containerd team reports, "these issues are resolved in containerd v1.6.4."

With containerd v1.6.0-v1.6.3, if you do not upgrade the CNI plugins and/or declare the CNI config version, you might encounter the following "Incompatible CNI versions" or "Failed to destroy network for sandbox" error conditions.

Incompatible CNI versions error

If the version of your CNI plugin does not correctly match the plugin version in the config because the config version is later than the plugin version, the containerd log will likely show an error message on startup of a pod similar to:

incompatible CNI versions; config is \"1.0.0\", plugin supports [\"0.1.0\" \"0.2.0\" \"0.3.0\" \"0.3.1\" \"0.4.0\"]"

To fix this issue, update your CNI plugins and CNI config files.

Failed to destroy network for sandbox error

If the version of the plugin is missing in the CNI plugin config, the pod may run. However, stopping the pod generates an error similar to:

ERROR[2022-04-26T00:43:24.518165483Z] StopPodSandbox for "b" failed
error="failed to destroy network for sandbox \"bbc85f891eaf060c5a879e27bba9b6b06450210161dfdecfbb2732959fb6500a\": invalid version \"\": the version is empty"

This error leaves the pod in the not-ready state with a network namespace still attached. To recover from this problem, edit the CNI config file to add the missing version information. The next attempt to stop the pod should be successful.

Updating your CNI plugins and CNI config files

If you're using containerd v1.6.0-v1.6.3 and encountered "Incompatible CNI versions" or "Failed to destroy network for sandbox" errors, consider updating your CNI plugins and editing the CNI config files.

Here's an overview of the typical steps for each node:

  1. Safely drain and cordon the node.
  2. After stopping your container runtime and kubelet services, perform the following upgrade operations:
  • If you're running CNI plugins, upgrade them to the latest version.
  • If you're using non-CNI plugins, replace them with CNI plugins. Use the latest version of the plugins.
  • Update the plugin configuration file to specify or match a version of the CNI specification that the plugin supports, as shown in the following "An example containerd configuration file" section.
  • For containerd, ensure that you have installed the latest version (v1.0.0 or later) of the CNI loopback plugin.
  • Upgrade node components (for example, the kubelet) to Kubernetes v1.24
  • Upgrade to or install the most current version of the container runtime.
  1. Bring the node back into your cluster by restarting your container runtime and kubelet. Uncordon the node (kubectl uncordon <nodename>).

An example containerd configuration file

The following example shows a configuration for containerd runtime v1.6.x, which supports a recent version of the CNI specification (v1.0.0).

Please see the documentation from your plugin and networking provider for further instructions on configuring your system.

On Kubernetes, containerd runtime adds a loopback interface, lo, to pods as a default behavior. The containerd runtime configures the loopback interface via a CNI plugin, loopback. The loopback plugin is distributed as part of the containerd release packages that have the cni designation. containerd v1.6.0 and later includes a CNI v1.0.0-compatible loopback plugin as well as other default CNI plugins. The configuration for the loopback plugin is done internally by containerd, and is set to use CNI v1.0.0. This also means that the version of the loopback plugin must be v1.0.0 or later when this newer version containerd is started.

The following bash command generates an example CNI config. Here, the 1.0.0 value for the config version is assigned to the cniVersion field for use when containerd invokes the CNI bridge plugin.

cat << EOF | tee /etc/cni/net.d/10-containerd-net.conflist
{
 "cniVersion": "1.0.0",
 "name": "containerd-net",
 "plugins": [
   {
     "type": "bridge",
     "bridge": "cni0",
     "isGateway": true,
     "ipMasq": true,
     "promiscMode": true,
     "ipam": {
       "type": "host-local",
       "ranges": [
         [{
           "subnet": "10.88.0.0/16"
         }],
         [{
           "subnet": "2001:db8:4860::/64"
         }]
       ],
       "routes": [
         { "dst": "0.0.0.0/0" },
         { "dst": "::/0" }
       ]
     }
   },
   {
     "type": "portmap",
     "capabilities": {"portMappings": true},
     "externalSetMarkChain": "KUBE-MARK-MASQ"
   }
 ]
}
EOF

Update the IP address ranges in the preceding example with ones that are based on your use case and network addressing plan.

3.5 - Check whether dockershim removal affects you

The dockershim component of Kubernetes allows the use of Docker as a Kubernetes's container runtime. Kubernetes' built-in dockershim component was removed in release v1.24.

This page explains how your cluster could be using Docker as a container runtime, provides details on the role that dockershim plays when in use, and shows steps you can take to check whether any workloads could be affected by dockershim removal.

Finding if your app has a dependencies on Docker

If you are using Docker for building your application containers, you can still run these containers on any container runtime. This use of Docker does not count as a dependency on Docker as a container runtime.

When alternative container runtime is used, executing Docker commands may either not work or yield unexpected output. This is how you can find whether you have a dependency on Docker:

  1. Make sure no privileged Pods execute Docker commands (like docker ps), restart the Docker service (commands such as systemctl restart docker.service), or modify Docker-specific files such as /etc/docker/daemon.json.
  2. Check for any private registries or image mirror settings in the Docker configuration file (like /etc/docker/daemon.json). Those typically need to be reconfigured for another container runtime.
  3. Check that scripts and apps running on nodes outside of your Kubernetes infrastructure do not execute Docker commands. It might be:
    • SSH to nodes to troubleshoot;
    • Node startup scripts;
    • Monitoring and security agents installed on nodes directly.
  4. Third-party tools that perform above mentioned privileged operations. See Migrating telemetry and security agents from dockershim for more information.
  5. Make sure there are no indirect dependencies on dockershim behavior. This is an edge case and unlikely to affect your application. Some tooling may be configured to react to Docker-specific behaviors, for example, raise alert on specific metrics or search for a specific log message as part of troubleshooting instructions. If you have such tooling configured, test the behavior on a test cluster before migration.

Dependency on Docker explained

A container runtime is software that can execute the containers that make up a Kubernetes pod. Kubernetes is responsible for orchestration and scheduling of Pods; on each node, the kubelet uses the container runtime interface as an abstraction so that you can use any compatible container runtime.

In its earliest releases, Kubernetes offered compatibility with one container runtime: Docker. Later in the Kubernetes project's history, cluster operators wanted to adopt additional container runtimes. The CRI was designed to allow this kind of flexibility - and the kubelet began supporting CRI. However, because Docker existed before the CRI specification was invented, the Kubernetes project created an adapter component, dockershim. The dockershim adapter allows the kubelet to interact with Docker as if Docker were a CRI compatible runtime.

You can read about it in Kubernetes Containerd integration goes GA blog post.

Dockershim vs. CRI with Containerd

Switching to Containerd as a container runtime eliminates the middleman. All the same containers can be run by container runtimes like Containerd as before. But now, since containers schedule directly with the container runtime, they are not visible to Docker. So any Docker tooling or fancy UI you might have used before to check on these containers is no longer available.

You cannot get container information using docker ps or docker inspect commands. As you cannot list containers, you cannot get logs, stop containers, or execute something inside a container using docker exec.

You can still pull images or build them using docker build command. But images built or pulled by Docker would not be visible to container runtime and Kubernetes. They needed to be pushed to some registry to allow them to be used by Kubernetes.

Known issues

Some filesystem metrics are missing and the metrics format is different

The Kubelet /metrics/cadvisor endpoint provides Prometheus metrics, as documented in Metrics for Kubernetes system components. If you install a metrics collector that depends on that endpoint, you might see the following issues:

  • The metrics format on the Docker node is k8s_<container-name>_<pod-name>_<namespace>_<pod-uid>_<restart-count> but the format on other runtime is different. For example, on containerd node it is <container-id>.
  • Some filesystem metrics are missing, as follows: 
    container_fs_inodes_free
container_fs_inodes_total
container_fs_io_current
container_fs_io_time_seconds_total
container_fs_io_time_weighted_seconds_total
container_fs_limit_bytes
container_fs_read_seconds_total
container_fs_reads_merged_total
container_fs_sector_reads_total
container_fs_sector_writes_total
container_fs_usage_bytes
container_fs_write_seconds_total
container_fs_writes_merged_total

Workaround

You can mitigate this issue by using cAdvisor as a standalone daemonset.

  1. Find the latest cAdvisor release with the name pattern vX.Y.Z-containerd-cri (for example, v0.42.0-containerd-cri).
  2. Follow the steps in cAdvisor Kubernetes Daemonset to create the daemonset.
  3. Point the installed metrics collector to use the cAdvisor /metrics endpoint which provides the full set of Prometheus container metrics.

Alternatives:

  • Use alternative third party metrics collection solution.
  • Collect metrics from the Kubelet summary API that is served at /stats/summary.

What's next

3.6 - Migrating telemetry and security agents from dockershim

Kubernetes' support for direct integration with Docker Engine is deprecated and has been removed. Most apps do not have a direct dependency on runtime hosting containers. However, there are still a lot of telemetry and monitoring agents that have a dependency on Docker to collect containers metadata, logs, and metrics. This document aggregates information on how to detect these dependencies as well as links on how to migrate these agents to use generic tools or alternative runtimes.

Telemetry and security agents

Within a Kubernetes cluster there are a few different ways to run telemetry or security agents. Some agents have a direct dependency on Docker Engine when they run as DaemonSets or directly on nodes.

Why do some telemetry agents communicate with Docker Engine?

Historically, Kubernetes was written to work specifically with Docker Engine. Kubernetes took care of networking and scheduling, relying on Docker Engine for launching and running containers (within Pods) on a node. Some information that is relevant to telemetry, such as a pod name, is only available from Kubernetes components. Other data, such as container metrics, is not the responsibility of the container runtime. Early telemetry agents needed to query the container runtime and Kubernetes to report an accurate picture. Over time, Kubernetes gained the ability to support multiple runtimes, and now supports any runtime that is compatible with the container runtime interface.

Some telemetry agents rely specifically on Docker Engine tooling. For example, an agent might run a command such as docker ps or docker top to list containers and processes or docker logs to receive streamed logs. If nodes in your existing cluster use Docker Engine, and you switch to a different container runtime, these commands will not work any longer.

Identify DaemonSets that depend on Docker Engine

If a pod wants to make calls to the dockerd running on the node, the pod must either:

  • mount the filesystem containing the Docker daemon's privileged socket, as a volume; or
  • mount the specific path of the Docker daemon's privileged socket directly, also as a volume.

For example: on COS images, Docker exposes its Unix domain socket at /var/run/docker.sock This means that the pod spec will include a hostPath volume mount of /var/run/docker.sock.

Here's a sample shell script to find Pods that have a mount directly mapping the Docker socket. This script outputs the namespace and name of the pod. You can remove the grep '/var/run/docker.sock' to review other mounts.

kubectl get pods --all-namespaces \
-o=jsonpath='{range .items[*]}{"\n"}{.metadata.namespace}{":\t"}{.metadata.name}{":\t"}{range .spec.volumes[*]}{.hostPath.path}{", "}{end}{end}' \
| sort \
| grep '/var/run/docker.sock'

Detecting Docker dependency from node agents

If your cluster nodes are customized and install additional security and telemetry agents on the node, check with the agent vendor to verify whether it has any dependency on Docker.

Telemetry and security agent vendors

This section is intended to aggregate information about various telemetry and security agents that may have a dependency on container runtimes.

We keep the work in progress version of migration instructions for various telemetry and security agent vendors in Google doc. Please contact the vendor to get up to date instructions for migrating from dockershim.

Migration from dockershim

Aqua

No changes are needed: everything should work seamlessly on the runtime switch.

Datadog

How to migrate: Docker deprecation in Kubernetes The pod that accesses Docker Engine may have a name containing any of:

  • datadog-agent
  • datadog
  • dd-agent

Dynatrace

How to migrate: Migrating from Docker-only to generic container metrics in Dynatrace

Containerd support announcement: Get automated full-stack visibility into containerd-based Kubernetes environments

CRI-O support announcement: Get automated full-stack visibility into your CRI-O Kubernetes containers (Beta)

The pod accessing Docker may have name containing:

  • dynatrace-oneagent

Falco

How to migrate:

Migrate Falco from dockershim Falco supports any CRI-compatible runtime (containerd is used in the default configuration); the documentation explains all details. The pod accessing Docker may have name containing:

  • falco

Prisma Cloud Compute

Check documentation for Prisma Cloud, under the "Install Prisma Cloud on a CRI (non-Docker) cluster" section. The pod accessing Docker may be named like:

  • twistlock-defender-ds

SignalFx (Splunk)

The SignalFx Smart Agent (deprecated) uses several different monitors for Kubernetes including kubernetes-cluster, kubelet-stats/kubelet-metrics, and docker-container-stats. The kubelet-stats monitor was previously deprecated by the vendor, in favor of kubelet-metrics. The docker-container-stats monitor is the one affected by dockershim removal. Do not use the docker-container-stats with container runtimes other than Docker Engine.

How to migrate from dockershim-dependent agent:

  1. Remove docker-container-stats from the list of configured monitors. Note, keeping this monitor enabled with non-dockershim runtime will result in incorrect metrics being reported when docker is installed on node and no metrics when docker is not installed.
  2. Enable and configure kubelet-metrics monitor.

The Pod accessing Docker may be named something like:

  • signalfx-agent

Yahoo Kubectl Flame

Flame does not support container runtimes other than Docker. See https://github.com/yahoo/kubectl-flame/issues/51

4 - Generate Certificates Manually

When using client certificate authentication, you can generate certificates manually through easyrsa, openssl or cfssl.

easyrsa

easyrsa can manually generate certificates for your cluster.

  1. Download, unpack, and initialize the patched version of easyrsa3.

    curl -LO https://dl.k8s.io/easy-rsa/easy-rsa.tar.gz
tar xzf easy-rsa.tar.gz
cd easy-rsa-master/easyrsa3
./easyrsa init-pki

  2. Generate a new certificate authority (CA). --batch sets automatic mode; --req-cn specifies the Common Name (CN) for the CA's new root certificate.

    ./easyrsa --batch "--req-cn=${MASTER_IP}@`date +%s`" build-ca nopass

  3. Generate server certificate and key.

    The argument --subject-alt-name sets the possible IPs and DNS names the API server will be accessed with. The MASTER_CLUSTER_IP is usually the first IP from the service CIDR that is specified as the --service-cluster-ip-range argument for both the API server and the controller manager component. The argument --days is used to set the number of days after which the certificate expires. The sample below also assumes that you are using cluster.local as the default DNS domain name.

    ./easyrsa --subject-alt-name="IP:${MASTER_IP},"\
"IP:${MASTER_CLUSTER_IP},"\
"DNS:kubernetes,"\
"DNS:kubernetes.default,"\
"DNS:kubernetes.default.svc,"\
"DNS:kubernetes.default.svc.cluster,"\
"DNS:kubernetes.default.svc.cluster.local" \
--days=10000 \
build-server-full server nopass

  4. Copy pki/ca.crt, pki/issued/server.crt, and pki/private/server.key to your directory.

  5. Fill in and add the following parameters into the API server start parameters:

    --client-ca-file=/yourdirectory/ca.crt
--tls-cert-file=/yourdirectory/server.crt
--tls-private-key-file=/yourdirectory/server.key

openssl

openssl can manually generate certificates for your cluster.

  1. Generate a ca.key with 2048bit:

    openssl genrsa -out ca.key 2048

  2. According to the ca.key generate a ca.crt (use -days to set the certificate effective time):

    openssl req -x509 -new -nodes -key ca.key -subj "/CN=${MASTER_IP}" -days 10000 -out ca.crt

  3. Generate a server.key with 2048bit:

    openssl genrsa -out server.key 2048

  4. Create a config file for generating a Certificate Signing Request (CSR).

    Be sure to substitute the values marked with angle brackets (e.g. <MASTER_IP>) with real values before saving this to a file (e.g. csr.conf). Note that the value for MASTER_CLUSTER_IP is the service cluster IP for the API server as described in previous subsection. The sample below also assumes that you are using cluster.local as the default DNS domain name.

    [ req ]
default_bits = 2048
prompt = no
default_md = sha256
req_extensions = req_ext
distinguished_name = dn

[ dn ]
C = <country>
ST = <state>
L = <city>
O = <organization>
OU = <organization unit>
CN = <MASTER_IP>

[ req_ext ]
subjectAltName = @alt_names

[ alt_names ]
DNS.1 = kubernetes
DNS.2 = kubernetes.default
DNS.3 = kubernetes.default.svc
DNS.4 = kubernetes.default.svc.cluster
DNS.5 = kubernetes.default.svc.cluster.local
IP.1 = <MASTER_IP>
IP.2 = <MASTER_CLUSTER_IP>

[ v3_ext ]
authorityKeyIdentifier=keyid,issuer:always
basicConstraints=CA:FALSE
keyUsage=keyEncipherment,dataEncipherment
extendedKeyUsage=serverAuth,clientAuth
subjectAltName=@alt_names

  5. Generate the certificate signing request based on the config file:

    openssl req -new -key server.key -out server.csr -config csr.conf

  6. Generate the server certificate using the ca.key, ca.crt and server.csr:

    openssl x509 -req -in server.csr -CA ca.crt -CAkey ca.key \
    -CAcreateserial -out server.crt -days 10000 \
    -extensions v3_ext -extfile csr.conf -sha256

  7. View the certificate signing request:

    openssl req  -noout -text -in ./server.csr

  8. View the certificate:

    openssl x509  -noout -text -in ./server.crt

Finally, add the same parameters into the API server start parameters.

cfssl

cfssl is another tool for certificate generation.

  1. Download, unpack and prepare the command line tools as shown below.

    Note that you may need to adapt the sample commands based on the hardware architecture and cfssl version you are using.

    curl -L https://github.com/cloudflare/cfssl/releases/download/v1.5.0/cfssl_1.5.0_linux_amd64 -o cfssl
chmod +x cfssl
curl -L https://github.com/cloudflare/cfssl/releases/download/v1.5.0/cfssljson_1.5.0_linux_amd64 -o cfssljson
chmod +x cfssljson
curl -L https://github.com/cloudflare/cfssl/releases/download/v1.5.0/cfssl-certinfo_1.5.0_linux_amd64 -o cfssl-certinfo
chmod +x cfssl-certinfo

  2. Create a directory to hold the artifacts and initialize cfssl:

    mkdir cert
cd cert
../cfssl print-defaults config > config.json
../cfssl print-defaults csr > csr.json

  3. Create a JSON config file for generating the CA file, for example, ca-config.json:

    {
  "signing": {
    "default": {
      "expiry": "8760h"
    },
    "profiles": {
      "kubernetes": {
        "usages": [
          "signing",
          "key encipherment",
          "server auth",
          "client auth"
        ],
        "expiry": "8760h"
      }
    }
  }
}

  4. Create a JSON config file for CA certificate signing request (CSR), for example, ca-csr.json. Be sure to replace the values marked with angle brackets with real values you want to use.

    {
  "CN": "kubernetes",
  "key": {
    "algo": "rsa",
    "size": 2048
  },
  "names":[{
    "C": "<country>",
    "ST": "<state>",
    "L": "<city>",
    "O": "<organization>",
    "OU": "<organization unit>"
  }]
}

  5. Generate CA key (ca-key.pem) and certificate (ca.pem):

    ../cfssl gencert -initca ca-csr.json | ../cfssljson -bare ca

  6. Create a JSON config file for generating keys and certificates for the API server, for example, server-csr.json. Be sure to replace the values in angle brackets with real values you want to use. The <MASTER_CLUSTER_IP> is the service cluster IP for the API server as described in previous subsection. The sample below also assumes that you are using cluster.local as the default DNS domain name.

    {
  "CN": "kubernetes",
  "hosts": [
    "127.0.0.1",
    "<MASTER_IP>",
    "<MASTER_CLUSTER_IP>",
    "kubernetes",
    "kubernetes.default",
    "kubernetes.default.svc",
    "kubernetes.default.svc.cluster",
    "kubernetes.default.svc.cluster.local"
  ],
  "key": {
    "algo": "rsa",
    "size": 2048
  },
  "names": [{
    "C": "<country>",
    "ST": "<state>",
    "L": "<city>",
    "O": "<organization>",
    "OU": "<organization unit>"
  }]
}

  7. Generate the key and certificate for the API server, which are by default saved into file server-key.pem and server.pem respectively:

    ../cfssl gencert -ca=ca.pem -ca-key=ca-key.pem \
     --config=ca-config.json -profile=kubernetes \
     server-csr.json | ../cfssljson -bare server

Distributing Self-Signed CA Certificate

A client node may refuse to recognize a self-signed CA certificate as valid. For a non-production deployment, or for a deployment that runs behind a company firewall, you can distribute a self-signed CA certificate to all clients and refresh the local list for valid certificates.

On each client, perform the following operations:

sudo cp ca.crt /usr/local/share/ca-certificates/kubernetes.crt
sudo update-ca-certificates

Updating certificates in /etc/ssl/certs...
1 added, 0 removed; done.
Running hooks in /etc/ca-certificates/update.d....
done.

Certificates API

You can use the certificates.k8s.io API to provision x509 certificates to use for authentication as documented in the Managing TLS in a cluster task page.

5 - Manage Memory, CPU, and API Resources

5.1 - Configure Default Memory Requests and Limits for a Namespace

Define a default memory resource limit for a namespace, so that every new Pod in that namespace has a memory resource limit configured.

This page shows how to configure default memory requests and limits for a namespace.

A Kubernetes cluster can be divided into namespaces. Once you have a namespace that has a default memory limit, and you then try to create a Pod with a container that does not specify its own memory limit, then the control plane assigns the default memory limit to that container.

Kubernetes assigns a default memory request under certain conditions that are explained later in this topic.

Before you begin

You need to have a Kubernetes cluster, and the kubectl command-line tool must be configured to communicate with your cluster. It is recommended to run this tutorial on a cluster with at least two nodes that are not acting as control plane hosts. If you do not already have a cluster, you can create one by using minikube or you can use one of these Kubernetes playgrounds:

You must have access to create namespaces in your cluster.

Each node in your cluster must have at least 2 GiB of memory.

Create a namespace

Create a namespace so that the resources you create in this exercise are isolated from the rest of your cluster.

kubectl create namespace default-mem-example

Create a LimitRange and a Pod

Here's a manifest for an example LimitRange. The manifest specifies a default memory request and a default memory limit.

admin/resource/memory-defaults.yaml 
apiVersion: v1
kind: LimitRange
metadata:
  name: mem-limit-range
spec:
  limits:
  - default:
      memory: 512Mi
    defaultRequest:
      memory: 256Mi
    type: Container

Create the LimitRange in the default-mem-example namespace:

kubectl apply -f https://k8s.io/examples/admin/resource/memory-defaults.yaml --namespace=default-mem-example

Now if you create a Pod in the default-mem-example namespace, and any container within that Pod does not specify its own values for memory request and memory limit, then the control plane applies default values: a memory request of 256MiB and a memory limit of 512MiB.

Here's an example manifest for a Pod that has one container. The container does not specify a memory request and limit.

admin/resource/memory-defaults-pod.yaml 
apiVersion: v1
kind: Pod
metadata:
  name: default-mem-demo
spec:
  containers:
  - name: default-mem-demo-ctr
    image: nginx

Create the Pod.

kubectl apply -f https://k8s.io/examples/admin/resource/memory-defaults-pod.yaml --namespace=default-mem-example

View detailed information about the Pod:

kubectl get pod default-mem-demo --output=yaml --namespace=default-mem-example

The output shows that the Pod's container has a memory request of 256 MiB and a memory limit of 512 MiB. These are the default values specified by the LimitRange.

containers:
- image: nginx
  imagePullPolicy: Always
  name: default-mem-demo-ctr
  resources:
    limits:
      memory: 512Mi
    requests:
      memory: 256Mi

Delete your Pod:

kubectl delete pod default-mem-demo --namespace=default-mem-example

What if you specify a container's limit, but not its request?

Here's a manifest for a Pod that has one container. The container specifies a memory limit, but not a request:

admin/resource/memory-defaults-pod-2.yaml 
apiVersion: v1
kind: Pod
metadata:
  name: default-mem-demo-2
spec:
  containers:
  - name: default-mem-demo-2-ctr
    image: nginx
    resources:
      limits:
        memory: "1Gi"

Create the Pod:

kubectl apply -f https://k8s.io/examples/admin/resource/memory-defaults-pod-2.yaml --namespace=default-mem-example

View detailed information about the Pod:

kubectl get pod default-mem-demo-2 --output=yaml --namespace=default-mem-example

The output shows that the container's memory request is set to match its memory limit. Notice that the container was not assigned the default memory request value of 256Mi.

resources:
  limits:
    memory: 1Gi
  requests:
    memory: 1Gi

What if you specify a container's request, but not its limit?

Here's a manifest for a Pod that has one container. The container specifies a memory request, but not a limit:

admin/resource/memory-defaults-pod-3.yaml 
apiVersion: v1
kind: Pod
metadata:
  name: default-mem-demo-3
spec:
  containers:
  - name: default-mem-demo-3-ctr
    image: nginx
    resources:
      requests:
        memory: "128Mi"

Create the Pod:

kubectl apply -f https://k8s.io/examples/admin/resource/memory-defaults-pod-3.yaml --namespace=default-mem-example

View the Pod's specification:

kubectl get pod default-mem-demo-3 --output=yaml --namespace=default-mem-example

The output shows that the container's memory request is set to the value specified in the container's manifest. The container is limited to use no more than 512MiB of memory, which matches the default memory limit for the namespace.

resources:
  limits:
    memory: 512Mi
  requests:
    memory: 128Mi

Motivation for default memory limits and requests

If your namespace has a memory resource quota configured, it is helpful to have a default value in place for memory limit. Here are three of the restrictions that a resource quota imposes on a namespace:

  • For every Pod that runs in the namespace, the Pod and each of its containers must have a memory limit. (If you specify a memory limit for every container in a Pod, Kubernetes can infer the Pod-level memory limit by adding up the limits for its containers).
  • Memory limits apply a resource reservation on the node where the Pod in question is scheduled. The total amount of memory reserved for all Pods in the namespace must not exceed a specified limit.
  • The total amount of memory actually used by all Pods in the namespace must also not exceed a specified limit.

When you add a LimitRange:

If any Pod in that namespace that includes a container does not specify its own memory limit, the control plane applies the default memory limit to that container, and the Pod can be allowed to run in a namespace that is restricted by a memory ResourceQuota.

Clean up

Delete your namespace:

kubectl delete namespace default-mem-example

What's next

For cluster administrators

For app developers

5.2 - Configure Default CPU Requests and Limits for a Namespace

Define a default CPU resource limits for a namespace, so that every new Pod in that namespace has a CPU resource limit configured.

This page shows how to configure default CPU requests and limits for a namespace.

A Kubernetes cluster can be divided into namespaces. If you create a Pod within a namespace that has a default CPU limit, and any container in that Pod does not specify its own CPU limit, then the control plane assigns the default CPU limit to that container.

Kubernetes assigns a default CPU request, but only under certain conditions that are explained later in this page.

Before you begin

You need to have a Kubernetes cluster, and the kubectl command-line tool must be configured to communicate with your cluster. It is recommended to run this tutorial on a cluster with at least two nodes that are not acting as control plane hosts. If you do not already have a cluster, you can create one by using minikube or you can use one of these Kubernetes playgrounds:

You must have access to create namespaces in your cluster.

If you're not already familiar with what Kubernetes means by 1.0 CPU, read meaning of CPU.

Create a namespace

Create a namespace so that the resources you create in this exercise are isolated from the rest of your cluster.

kubectl create namespace default-cpu-example

Create a LimitRange and a Pod

Here's a manifest for an example LimitRange. The manifest specifies a default CPU request and a default CPU limit.

admin/resource/cpu-defaults.yaml 
apiVersion: v1
kind: LimitRange
metadata:
  name: cpu-limit-range
spec:
  limits:
  - default:
      cpu: 1
    defaultRequest:
      cpu: 0.5
    type: Container

Create the LimitRange in the default-cpu-example namespace:

kubectl apply -f https://k8s.io/examples/admin/resource/cpu-defaults.yaml --namespace=default-cpu-example

Now if you create a Pod in the default-cpu-example namespace, and any container in that Pod does not specify its own values for CPU request and CPU limit, then the control plane applies default values: a CPU request of 0.5 and a default CPU limit of 1.

Here's a manifest for a Pod that has one container. The container does not specify a CPU request and limit.

admin/resource/cpu-defaults-pod.yaml 
apiVersion: v1
kind: Pod
metadata:
  name: default-cpu-demo
spec:
  containers:
  - name: default-cpu-demo-ctr
    image: nginx

Create the Pod.

kubectl apply -f https://k8s.io/examples/admin/resource/cpu-defaults-pod.yaml --namespace=default-cpu-example

View the Pod's specification:

kubectl get pod default-cpu-demo --output=yaml --namespace=default-cpu-example

The output shows that the Pod's only container has a CPU request of 500m cpu (which you can read as “500 millicpu”), and a CPU limit of 1 cpu. These are the default values specified by the LimitRange.

containers:
- image: nginx
  imagePullPolicy: Always
  name: default-cpu-demo-ctr
  resources:
    limits:
      cpu: "1"
    requests:
      cpu: 500m

What if you specify a container's limit, but not its request?

Here's a manifest for a Pod that has one container. The container specifies a CPU limit, but not a request:

admin/resource/cpu-defaults-pod-2.yaml 
apiVersion: v1
kind: Pod
metadata:
  name: default-cpu-demo-2
spec:
  containers:
  - name: default-cpu-demo-2-ctr
    image: nginx
    resources:
      limits:
        cpu: "1"

Create the Pod:

kubectl apply -f https://k8s.io/examples/admin/resource/cpu-defaults-pod-2.yaml --namespace=default-cpu-example

View the specification of the Pod that you created:

kubectl get pod default-cpu-demo-2 --output=yaml --namespace=default-cpu-example

The output shows that the container's CPU request is set to match its CPU limit. Notice that the container was not assigned the default CPU request value of 0.5 cpu:

resources:
  limits:
    cpu: "1"
  requests:
    cpu: "1"

What if you specify a container's request, but not its limit?

Here's an example manifest for a Pod that has one container. The container specifies a CPU request, but not a limit:

admin/resource/cpu-defaults-pod-3.yaml 
apiVersion: v1
kind: Pod
metadata:
  name: default-cpu-demo-3
spec:
  containers:
  - name: default-cpu-demo-3-ctr
    image: nginx
    resources:
      requests:
        cpu: "0.75"

Create the Pod:

kubectl apply -f https://k8s.io/examples/admin/resource/cpu-defaults-pod-3.yaml --namespace=default-cpu-example

View the specification of the Pod that you created:

kubectl get pod default-cpu-demo-3 --output=yaml --namespace=default-cpu-example

The output shows that the container's CPU request is set to the value you specified at the time you created the Pod (in other words: it matches the manifest). However, the same container's CPU limit is set to 1 cpu, which is the default CPU limit for that namespace.

resources:
  limits:
    cpu: "1"
  requests:
    cpu: 750m

Motivation for default CPU limits and requests

If your namespace has a CPU resource quota configured, it is helpful to have a default value in place for CPU limit. Here are two of the restrictions that a CPU resource quota imposes on a namespace:

  • For every Pod that runs in the namespace, each of its containers must have a CPU limit.
  • CPU limits apply a resource reservation on the node where the Pod in question is scheduled. The total amount of CPU that is reserved for use by all Pods in the namespace must not exceed a specified limit.

When you add a LimitRange:

If any Pod in that namespace that includes a container does not specify its own CPU limit, the control plane applies the default CPU limit to that container, and the Pod can be allowed to run in a namespace that is restricted by a CPU ResourceQuota.

Clean up

Delete your namespace:

kubectl delete namespace default-cpu-example

What's next

For cluster administrators

For app developers

5.3 - Configure Minimum and Maximum Memory Constraints for a Namespace

Define a range of valid memory resource limits for a namespace, so that every new Pod in that namespace falls within the range you configure.

This page shows how to set minimum and maximum values for memory used by containers running in a namespace. You specify minimum and maximum memory values in a LimitRange object. If a Pod does not meet the constraints imposed by the LimitRange, it cannot be created in the namespace.

Before you begin

You need to have a Kubernetes cluster, and the kubectl command-line tool must be configured to communicate with your cluster. It is recommended to run this tutorial on a cluster with at least two nodes that are not acting as control plane hosts. If you do not already have a cluster, you can create one by using minikube or you can use one of these Kubernetes playgrounds:

You must have access to create namespaces in your cluster.

Each node in your cluster must have at least 1 GiB of memory available for Pods.

Create a namespace

Create a namespace so that the resources you create in this exercise are isolated from the rest of your cluster.

kubectl create namespace constraints-mem-example

Create a LimitRange and a Pod

Here's an example manifest for a LimitRange:

admin/resource/memory-constraints.yaml 
apiVersion: v1
kind: LimitRange
metadata:
  name: mem-min-max-demo-lr
spec:
  limits:
  - max:
      memory: 1Gi
    min:
      memory: 500Mi
    type: Container

Create the LimitRange:

kubectl apply -f https://k8s.io/examples/admin/resource/memory-constraints.yaml --namespace=constraints-mem-example

View detailed information about the LimitRange:

kubectl get limitrange mem-min-max-demo-lr --namespace=constraints-mem-example --output=yaml

The output shows the minimum and maximum memory constraints as expected. But notice that even though you didn't specify default values in the configuration file for the LimitRange, they were created automatically.

  limits:
  - default:
      memory: 1Gi
    defaultRequest:
      memory: 1Gi
    max:
      memory: 1Gi
    min:
      memory: 500Mi
    type: Container

Now whenever you define a Pod within the constraints-mem-example namespace, Kubernetes performs these steps:

  • If any container in that Pod does not specify its own memory request and limit, the control plane assigns the default memory request and limit to that container.

  • Verify that every container in that Pod requests at least 500 MiB of memory.

  • Verify that every container in that Pod requests no more than 1024 MiB (1 GiB) of memory.

Here's a manifest for a Pod that has one container. Within the Pod spec, the sole container specifies a memory request of 600 MiB and a memory limit of 800 MiB. These satisfy the minimum and maximum memory constraints imposed by the LimitRange.

admin/resource/memory-constraints-pod.yaml 
apiVersion: v1
kind: Pod
metadata:
  name: constraints-mem-demo
spec:
  containers:
  - name: constraints-mem-demo-ctr
    image: nginx
    resources:
      limits:
        memory: "800Mi"
      requests:
        memory: "600Mi"

Create the Pod:

kubectl apply -f https://k8s.io/examples/admin/resource/memory-constraints-pod.yaml --namespace=constraints-mem-example

Verify that the Pod is running and that its container is healthy:

kubectl get pod constraints-mem-demo --namespace=constraints-mem-example

View detailed information about the Pod:

kubectl get pod constraints-mem-demo --output=yaml --namespace=constraints-mem-example

The output shows that the container within that Pod has a memory request of 600 MiB and a memory limit of 800 MiB. These satisfy the constraints imposed by the LimitRange for this namespace:

resources:
  limits:
     memory: 800Mi
  requests:
    memory: 600Mi

Delete your Pod:

kubectl delete pod constraints-mem-demo --namespace=constraints-mem-example

Attempt to create a Pod that exceeds the maximum memory constraint

Here's a manifest for a Pod that has one container. The container specifies a memory request of 800 MiB and a memory limit of 1.5 GiB.

admin/resource/memory-constraints-pod-2.yaml 
apiVersion: v1
kind: Pod
metadata:
  name: constraints-mem-demo-2
spec:
  containers:
  - name: constraints-mem-demo-2-ctr
    image: nginx
    resources:
      limits:
        memory: "1.5Gi"
      requests:
        memory: "800Mi"

Attempt to create the Pod:

kubectl apply -f https://k8s.io/examples/admin/resource/memory-constraints-pod-2.yaml --namespace=constraints-mem-example

The output shows that the Pod does not get created, because it defines a container that requests more memory than is allowed:

Error from server (Forbidden): error when creating "examples/admin/resource/memory-constraints-pod-2.yaml":
pods "constraints-mem-demo-2" is forbidden: maximum memory usage per Container is 1Gi, but limit is 1536Mi.

Attempt to create a Pod that does not meet the minimum memory request

Here's a manifest for a Pod that has one container. That container specifies a memory request of 100 MiB and a memory limit of 800 MiB.

admin/resource/memory-constraints-pod-3.yaml 
apiVersion: v1
kind: Pod
metadata:
  name: constraints-mem-demo-3
spec:
  containers:
  - name: constraints-mem-demo-3-ctr
    image: nginx
    resources:
      limits:
        memory: "800Mi"
      requests:
        memory: "100Mi"

Attempt to create the Pod:

kubectl apply -f https://k8s.io/examples/admin/resource/memory-constraints-pod-3.yaml --namespace=constraints-mem-example

The output shows that the Pod does not get created, because it defines a container that requests less memory than the enforced minimum:

Error from server (Forbidden): error when creating "examples/admin/resource/memory-constraints-pod-3.yaml":
pods "constraints-mem-demo-3" is forbidden: minimum memory usage per Container is 500Mi, but request is 100Mi.

Create a Pod that does not specify any memory request or limit

Here's a manifest for a Pod that has one container. The container does not specify a memory request, and it does not specify a memory limit.

admin/resource/memory-constraints-pod-4.yaml 
apiVersion: v1
kind: Pod
metadata:
  name: constraints-mem-demo-4
spec:
  containers:
  - name: constraints-mem-demo-4-ctr
    image: nginx

Create the Pod:

kubectl apply -f https://k8s.io/examples/admin/resource/memory-constraints-pod-4.yaml --namespace=constraints-mem-example

View detailed information about the Pod:

kubectl get pod constraints-mem-demo-4 --namespace=constraints-mem-example --output=yaml

The output shows that the Pod's only container has a memory request of 1 GiB and a memory limit of 1 GiB. How did that container get those values?

resources:
  limits:
    memory: 1Gi
  requests:
    memory: 1Gi

Because your Pod did not define any memory request and limit for that container, the cluster applied a default memory request and limit from the LimitRange.

This means that the definition of that Pod shows those values. You can check it using kubectl describe:

# Look for the "Requests:" section of the output
kubectl describe pod constraints-mem-demo-4 --namespace=constraints-mem-example

At this point, your Pod might be running or it might not be running. Recall that a prerequisite for this task is that your Nodes have at least 1 GiB of memory. If each of your Nodes has only 1 GiB of memory, then there is not enough allocatable memory on any Node to accommodate a memory request of 1 GiB. If you happen to be using Nodes with 2 GiB of memory, then you probably have enough space to accommodate the 1 GiB request.

Delete your Pod:

kubectl delete pod constraints-mem-demo-4 --namespace=constraints-mem-example

Enforcement of minimum and maximum memory constraints

The maximum and minimum memory constraints imposed on a namespace by a LimitRange are enforced only when a Pod is created or updated. If you change the LimitRange, it does not affect Pods that were created previously.

Motivation for minimum and maximum memory constraints

As a cluster administrator, you might want to impose restrictions on the amount of memory that Pods can use. For example:

  • Each Node in a cluster has 2 GiB of memory. You do not want to accept any Pod that requests more than 2 GiB of memory, because no Node in the cluster can support the request.

  • A cluster is shared by your production and development departments. You want to allow production workloads to consume up to 8 GiB of memory, but you want development workloads to be limited to 512 MiB. You create separate namespaces for production and development, and you apply memory constraints to each namespace.

Clean up

Delete your namespace:

kubectl delete namespace constraints-mem-example

What's next

For cluster administrators

For app developers

5.4 - Configure Minimum and Maximum CPU Constraints for a Namespace

Define a range of valid CPU resource limits for a namespace, so that every new Pod in that namespace falls within the range you configure.

This page shows how to set minimum and maximum values for the CPU resources used by containers and Pods in a namespace. You specify minimum and maximum CPU values in a LimitRange object. If a Pod does not meet the constraints imposed by the LimitRange, it cannot be created in the namespace.

Before you begin

You need to have a Kubernetes cluster, and the kubectl command-line tool must be configured to communicate with your cluster. It is recommended to run this tutorial on a cluster with at least two nodes that are not acting as control plane hosts. If you do not already have a cluster, you can create one by using minikube or you can use one of these Kubernetes playgrounds:

You must have access to create namespaces in your cluster.

Each node in your cluster must have at least 1.0 CPU available for Pods. See meaning of CPU to learn what Kubernetes means by “1 CPU”.

Create a namespace

Create a namespace so that the resources you create in this exercise are isolated from the rest of your cluster.

kubectl create namespace constraints-cpu-example

Create a LimitRange and a Pod

Here's a manifest for an example LimitRange:

admin/resource/cpu-constraints.yaml 
apiVersion: v1
kind: LimitRange
metadata:
  name: cpu-min-max-demo-lr
spec:
  limits:
  - max:
      cpu: "800m"
    min:
      cpu: "200m"
    type: Container

Create the LimitRange:

kubectl apply -f https://k8s.io/examples/admin/resource/cpu-constraints.yaml --namespace=constraints-cpu-example

View detailed information about the LimitRange:

kubectl get limitrange cpu-min-max-demo-lr --output=yaml --namespace=constraints-cpu-example

The output shows the minimum and maximum CPU constraints as expected. But notice that even though you didn't specify default values in the configuration file for the LimitRange, they were created automatically.

limits:
- default:
    cpu: 800m
  defaultRequest:
    cpu: 800m
  max:
    cpu: 800m
  min:
    cpu: 200m
  type: Container

Now whenever you create a Pod in the constraints-cpu-example namespace (or some other client of the Kubernetes API creates an equivalent Pod), Kubernetes performs these steps:

  • If any container in that Pod does not specify its own CPU request and limit, the control plane assigns the default CPU request and limit to that container.

  • Verify that every container in that Pod specifies a CPU request that is greater than or equal to 200 millicpu.

  • Verify that every container in that Pod specifies a CPU limit that is less than or equal to 800 millicpu.

Here's a manifest for a Pod that has one container. The container manifest specifies a CPU request of 500 millicpu and a CPU limit of 800 millicpu. These satisfy the minimum and maximum CPU constraints imposed by the LimitRange for this namespace.

admin/resource/cpu-constraints-pod.yaml 
apiVersion: v1
kind: Pod
metadata:
  name: constraints-cpu-demo
spec:
  containers:
  - name: constraints-cpu-demo-ctr
    image: nginx
    resources:
      limits:
        cpu: "800m"
      requests:
        cpu: "500m"

Create the Pod:

kubectl apply -f https://k8s.io/examples/admin/resource/cpu-constraints-pod.yaml --namespace=constraints-cpu-example

Verify that the Pod is running and that its container is healthy:

kubectl get pod constraints-cpu-demo --namespace=constraints-cpu-example

View detailed information about the Pod:

kubectl get pod constraints-cpu-demo --output=yaml --namespace=constraints-cpu-example

The output shows that the Pod's only container has a CPU request of 500 millicpu and CPU limit of 800 millicpu. These satisfy the constraints imposed by the LimitRange.

resources:
  limits:
    cpu: 800m
  requests:
    cpu: 500m

Delete the Pod

kubectl delete pod constraints-cpu-demo --namespace=constraints-cpu-example

Attempt to create a Pod that exceeds the maximum CPU constraint

Here's a manifest for a Pod that has one container. The container specifies a CPU request of 500 millicpu and a cpu limit of 1.5 cpu.

admin/resource/cpu-constraints-pod-2.yaml 
apiVersion: v1
kind: Pod
metadata:
  name: constraints-cpu-demo-2
spec:
  containers:
  - name: constraints-cpu-demo-2-ctr
    image: nginx
    resources:
      limits:
        cpu: "1.5"
      requests:
        cpu: "500m"

Attempt to create the Pod:

kubectl apply -f https://k8s.io/examples/admin/resource/cpu-constraints-pod-2.yaml --namespace=constraints-cpu-example

The output shows that the Pod does not get created, because it defines an unacceptable container. That container is not acceptable because it specifies a CPU limit that is too large:

Error from server (Forbidden): error when creating "examples/admin/resource/cpu-constraints-pod-2.yaml":
pods "constraints-cpu-demo-2" is forbidden: maximum cpu usage per Container is 800m, but limit is 1500m.

Attempt to create a Pod that does not meet the minimum CPU request

Here's a manifest for a Pod that has one container. The container specifies a CPU request of 100 millicpu and a CPU limit of 800 millicpu.

admin/resource/cpu-constraints-pod-3.yaml 
apiVersion: v1
kind: Pod
metadata:
  name: constraints-cpu-demo-3
spec:
  containers:
  - name: constraints-cpu-demo-3-ctr
    image: nginx
    resources:
      limits:
        cpu: "800m"
      requests:
        cpu: "100m"

Attempt to create the Pod:

kubectl apply -f https://k8s.io/examples/admin/resource/cpu-constraints-pod-3.yaml --namespace=constraints-cpu-example

The output shows that the Pod does not get created, because it defines an unacceptable container. That container is not acceptable because it specifies a CPU request that is lower than the enforced minimum:

Error from server (Forbidden): error when creating "examples/admin/resource/cpu-constraints-pod-3.yaml":
pods "constraints-cpu-demo-3" is forbidden: minimum cpu usage per Container is 200m, but request is 100m.

Create a Pod that does not specify any CPU request or limit

Here's a manifest for a Pod that has one container. The container does not specify a CPU request, nor does it specify a CPU limit.

admin/resource/cpu-constraints-pod-4.yaml 
apiVersion: v1
kind: Pod
metadata:
  name: constraints-cpu-demo-4
spec:
  containers:
  - name: constraints-cpu-demo-4-ctr
    image: vish/stress

Create the Pod:

kubectl apply -f https://k8s.io/examples/admin/resource/cpu-constraints-pod-4.yaml --namespace=constraints-cpu-example

View detailed information about the Pod:

kubectl get pod constraints-cpu-demo-4 --namespace=constraints-cpu-example --output=yaml

The output shows that the Pod's single container has a CPU request of 800 millicpu and a CPU limit of 800 millicpu. How did that container get those values?

resources:
  limits:
    cpu: 800m
  requests:
    cpu: 800m

Because that container did not specify its own CPU request and limit, the control plane applied the default CPU request and limit from the LimitRange for this namespace.

At this point, your Pod may or may not be running. Recall that a prerequisite for this task is that your Nodes must have at least 1 CPU available for use. If each of your Nodes has only 1 CPU, then there might not be enough allocatable CPU on any Node to accommodate a request of 800 millicpu. If you happen to be using Nodes with 2 CPU, then you probably have enough CPU to accommodate the 800 millicpu request.

Delete your Pod:

kubectl delete pod constraints-cpu-demo-4 --namespace=constraints-cpu-example

Enforcement of minimum and maximum CPU constraints

The maximum and minimum CPU constraints imposed on a namespace by a LimitRange are enforced only when a Pod is created or updated. If you change the LimitRange, it does not affect Pods that were created previously.

Motivation for minimum and maximum CPU constraints

As a cluster administrator, you might want to impose restrictions on the CPU resources that Pods can use. For example:

  • Each Node in a cluster has 2 CPU. You do not want to accept any Pod that requests more than 2 CPU, because no Node in the cluster can support the request.

  • A cluster is shared by your production and development departments. You want to allow production workloads to consume up to 3 CPU, but you want development workloads to be limited to 1 CPU. You create separate namespaces for production and development, and you apply CPU constraints to each namespace.

Clean up

Delete your namespace:

kubectl delete namespace constraints-cpu-example

What's next

For cluster administrators

For app developers

5.5 - Configure Memory and CPU Quotas for a Namespace

Define overall memory and CPU resource limits for a namespace.

This page shows how to set quotas for the total amount memory and CPU that can be used by all Pods running in a namespace. You specify quotas in a ResourceQuota object.

Before you begin

You need to have a Kubernetes cluster, and the kubectl command-line tool must be configured to communicate with your cluster. It is recommended to run this tutorial on a cluster with at least two nodes that are not acting as control plane hosts. If you do not already have a cluster, you can create one by using minikube or you can use one of these Kubernetes playgrounds:

You must have access to create namespaces in your cluster.

Each node in your cluster must have at least 1 GiB of memory.

Create a namespace

Create a namespace so that the resources you create in this exercise are isolated from the rest of your cluster.

kubectl create namespace quota-mem-cpu-example

Create a ResourceQuota

Here is a manifest for an example ResourceQuota:

admin/resource/quota-mem-cpu.yaml 
apiVersion: v1
kind: ResourceQuota
metadata:
  name: mem-cpu-demo
spec:
  hard:
    requests.cpu: "1"
    requests.memory: 1Gi
    limits.cpu: "2"
    limits.memory: 2Gi

Create the ResourceQuota:

kubectl apply -f https://k8s.io/examples/admin/resource/quota-mem-cpu.yaml --namespace=quota-mem-cpu-example

View detailed information about the ResourceQuota:

kubectl get resourcequota mem-cpu-demo --namespace=quota-mem-cpu-example --output=yaml

The ResourceQuota places these requirements on the quota-mem-cpu-example namespace:

  • For every Pod in the namespace, each container must have a memory request, memory limit, cpu request, and cpu limit.
  • The memory request total for all Pods in that namespace must not exceed 1 GiB.
  • The memory limit total for all Pods in that namespace must not exceed 2 GiB.
  • The CPU request total for all Pods in that namespace must not exceed 1 cpu.
  • The CPU limit total for all Pods in that namespace must not exceed 2 cpu.

See meaning of CPU to learn what Kubernetes means by “1 CPU”.

Create a Pod

Here is a manifest for an example Pod:

admin/resource/quota-mem-cpu-pod.yaml 
apiVersion: v1
kind: Pod
metadata:
  name: quota-mem-cpu-demo
spec:
  containers:
  - name: quota-mem-cpu-demo-ctr
    image: nginx
    resources:
      limits:
        memory: "800Mi"
        cpu: "800m"
      requests:
        memory: "600Mi"
        cpu: "400m"

Create the Pod:

kubectl apply -f https://k8s.io/examples/admin/resource/quota-mem-cpu-pod.yaml --namespace=quota-mem-cpu-example

Verify that the Pod is running and that its (only) container is healthy:

kubectl get pod quota-mem-cpu-demo --namespace=quota-mem-cpu-example

Once again, view detailed information about the ResourceQuota:

kubectl get resourcequota mem-cpu-demo --namespace=quota-mem-cpu-example --output=yaml

The output shows the quota along with how much of the quota has been used. You can see that the memory and CPU requests and limits for your Pod do not exceed the quota.

status:
  hard:
    limits.cpu: "2"
    limits.memory: 2Gi
    requests.cpu: "1"
    requests.memory: 1Gi
  used:
    limits.cpu: 800m
    limits.memory: 800Mi
    requests.cpu: 400m
    requests.memory: 600Mi

If you have the jq tool, you can also query (using JSONPath) for just the used values, and pretty-print that that of the output. For example:

kubectl get resourcequota mem-cpu-demo --namespace=quota-mem-cpu-example -o jsonpath='{ .status.used }' | jq .

Attempt to create a second Pod

Here is a manifest for a second Pod:

admin/resource/quota-mem-cpu-pod-2.yaml 
apiVersion: v1
kind: Pod
metadata:
  name: quota-mem-cpu-demo-2
spec:
  containers:
  - name: quota-mem-cpu-demo-2-ctr
    image: redis
    resources:
      limits:
        memory: "1Gi"
        cpu: "800m"
      requests:
        memory: "700Mi"
        cpu: "400m"

In the manifest, you can see that the Pod has a memory request of 700 MiB. Notice that the sum of the used memory request and this new memory request exceeds the memory request quota: 600 MiB + 700 MiB > 1 GiB.

Attempt to create the Pod:

kubectl apply -f https://k8s.io/examples/admin/resource/quota-mem-cpu-pod-2.yaml --namespace=quota-mem-cpu-example

The second Pod does not get created. The output shows that creating the second Pod would cause the memory request total to exceed the memory request quota.

Error from server (Forbidden): error when creating "examples/admin/resource/quota-mem-cpu-pod-2.yaml":
pods "quota-mem-cpu-demo-2" is forbidden: exceeded quota: mem-cpu-demo,
requested: requests.memory=700Mi,used: requests.memory=600Mi, limited: requests.memory=1Gi

Discussion

As you have seen in this exercise, you can use a ResourceQuota to restrict the memory request total for all Pods running in a namespace. You can also restrict the totals for memory limit, cpu request, and cpu limit.

Instead of managing total resource use within a namespace, you might want to restrict individual Pods, or the containers in those Pods. To achieve that kind of limiting, use a LimitRange.

Clean up

Delete your namespace:

kubectl delete namespace quota-mem-cpu-example

What's next

For cluster administrators

For app developers

5.6 - Configure a Pod Quota for a Namespace

Restrict how many Pods you can create within a namespace.

This page shows how to set a quota for the total number of Pods that can run in a Namespace. You specify quotas in a ResourceQuota object.

Before you begin

You need to have a Kubernetes cluster, and the kubectl command-line tool must be configured to communicate with your cluster. It is recommended to run this tutorial on a cluster with at least two nodes that are not acting as control plane hosts. If you do not already have a cluster, you can create one by using minikube or you can use one of these Kubernetes playgrounds:

You must have access to create namespaces in your cluster.

Create a namespace

Create a namespace so that the resources you create in this exercise are isolated from the rest of your cluster.

kubectl create namespace quota-pod-example

Create a ResourceQuota

Here is an example manifest for a ResourceQuota:

admin/resource/quota-pod.yaml 
apiVersion: v1
kind: ResourceQuota
metadata:
  name: pod-demo
spec:
  hard:
    pods: "2"

Create the ResourceQuota:

kubectl apply -f https://k8s.io/examples/admin/resource/quota-pod.yaml --namespace=quota-pod-example

View detailed information about the ResourceQuota:

kubectl get resourcequota pod-demo --namespace=quota-pod-example --output=yaml

The output shows that the namespace has a quota of two Pods, and that currently there are no Pods; that is, none of the quota is used.

spec:
  hard:
    pods: "2"
status:
  hard:
    pods: "2"
  used:
    pods: "0"

Here is an example manifest for a Deployment:

admin/resource/quota-pod-deployment.yaml 
apiVersion: apps/v1
kind: Deployment
metadata:
  name: pod-quota-demo
spec:
  selector:
    matchLabels:
      purpose: quota-demo
  replicas: 3
  template:
    metadata:
      labels:
        purpose: quota-demo
    spec:
      containers:
      - name: pod-quota-demo
        image: nginx

In that manifest, replicas: 3 tells Kubernetes to attempt to create three new Pods, all running the same application.

Create the Deployment:

kubectl apply -f https://k8s.io/examples/admin/resource/quota-pod-deployment.yaml --namespace=quota-pod-example

View detailed information about the Deployment:

kubectl get deployment pod-quota-demo --namespace=quota-pod-example --output=yaml

The output shows that even though the Deployment specifies three replicas, only two Pods were created because of the quota you defined earlier:

spec:
  ...
  replicas: 3
...
status:
  availableReplicas: 2
...
lastUpdateTime: 2021-04-02T20:57:05Z
    message: 'unable to create pods: pods "pod-quota-demo-1650323038-" is forbidden:
      exceeded quota: pod-demo, requested: pods=1, used: pods=2, limited: pods=2'

Choice of resource

In this task you have defined a ResourceQuota that limited the total number of Pods, but you could also limit the total number of other kinds of object. For example, you might decide to limit how many CronJobs that can live in a single namespace.

Clean up

Delete your namespace:

kubectl delete namespace quota-pod-example

What's next

For cluster administrators

For app developers

6 - Install a Network Policy Provider

6.1 - Use Antrea for NetworkPolicy

This page shows how to install and use Antrea CNI plugin on Kubernetes. For background on Project Antrea, read the Introduction to Antrea.

Before you begin

You need to have a Kubernetes cluster. Follow the kubeadm getting started guide to bootstrap one.

Deploying Antrea with kubeadm

Follow Getting Started guide to deploy Antrea for kubeadm.

What's next

Once your cluster is running, you can follow the Declare Network Policy to try out Kubernetes NetworkPolicy.

6.2 - Use Calico for NetworkPolicy

This page shows a couple of quick ways to create a Calico cluster on Kubernetes.

Before you begin

Decide whether you want to deploy a cloud or local cluster.

Creating a Calico cluster with Google Kubernetes Engine (GKE)

Prerequisite: gcloud.

  1. To launch a GKE cluster with Calico, include the --enable-network-policy flag.

    Syntax

    gcloud container clusters create [CLUSTER_NAME] --enable-network-policy

    Example

    gcloud container clusters create my-calico-cluster --enable-network-policy

  2. To verify the deployment, use the following command.

    kubectl get pods --namespace=kube-system

    The Calico pods begin with calico. Check to make sure each one has a status of Running.

Creating a local Calico cluster with kubeadm

To get a local single-host Calico cluster in fifteen minutes using kubeadm, refer to the Calico Quickstart.

What's next

Once your cluster is running, you can follow the Declare Network Policy to try out Kubernetes NetworkPolicy.

6.3 - Use Cilium for NetworkPolicy

This page shows how to use Cilium for NetworkPolicy.

For background on Cilium, read the Introduction to Cilium.

Before you begin

You need to have a Kubernetes cluster, and the kubectl command-line tool must be configured to communicate with your cluster. It is recommended to run this tutorial on a cluster with at least two nodes that are not acting as control plane hosts. If you do not already have a cluster, you can create one by using minikube or you can use one of these Kubernetes playgrounds:

To check the version, enter kubectl version.

Deploying Cilium on Minikube for Basic Testing

To get familiar with Cilium easily you can follow the Cilium Kubernetes Getting Started Guide to perform a basic DaemonSet installation of Cilium in minikube.

To start minikube, minimal version required is >= v1.5.2, run the with the following arguments:

minikube version

minikube version: v1.5.2

minikube start --network-plugin=cni

For minikube you can install Cilium using its CLI tool. To do so, first download the latest version of the CLI with the following command:

curl -LO https://github.com/cilium/cilium-cli/releases/latest/download/cilium-linux-amd64.tar.gz

Then extract the downloaded file to your /usr/local/bin directory with the following command:

sudo tar xzvfC cilium-linux-amd64.tar.gz /usr/local/bin
rm cilium-linux-amd64.tar.gz

After running the above commands, you can now install Cilium with the following command:

cilium install

Cilium will then automatically detect the cluster configuration and create and install the appropriate components for a successful installation. The components are:

  • Certificate Authority (CA) in Secret cilium-ca and certificates for Hubble (Cilium's observability layer).
  • Service accounts.
  • Cluster roles.
  • ConfigMap.
  • Agent DaemonSet and an Operator Deployment.

After the installation, you can view the overall status of the Cilium deployment with the cilium status command. See the expected output of the status command here.

The remainder of the Getting Started Guide explains how to enforce both L3/L4 (i.e., IP address + port) security policies, as well as L7 (e.g., HTTP) security policies using an example application.

Deploying Cilium for Production Use

For detailed instructions around deploying Cilium for production, see: Cilium Kubernetes Installation Guide This documentation includes detailed requirements, instructions and example production DaemonSet files.

Understanding Cilium components

Deploying a cluster with Cilium adds Pods to the kube-system namespace. To see this list of Pods run:

kubectl get pods --namespace=kube-system -l k8s-app=cilium

You'll see a list of Pods similar to this:

NAME           READY   STATUS    RESTARTS   AGE
cilium-kkdhz   1/1     Running   0          3m23s
...

A cilium Pod runs on each node in your cluster and enforces network policy on the traffic to/from Pods on that node using Linux BPF.

What's next

Once your cluster is running, you can follow the Declare Network Policy to try out Kubernetes NetworkPolicy with Cilium. Have fun, and if you have questions, contact us using the Cilium Slack Channel.

6.4 - Use Kube-router for NetworkPolicy

This page shows how to use Kube-router for NetworkPolicy.

Before you begin

You need to have a Kubernetes cluster running. If you do not already have a cluster, you can create one by using any of the cluster installers like Kops, Bootkube, Kubeadm etc.

Installing Kube-router addon

The Kube-router Addon comes with a Network Policy Controller that watches Kubernetes API server for any NetworkPolicy and pods updated and configures iptables rules and ipsets to allow or block traffic as directed by the policies. Please follow the trying Kube-router with cluster installers guide to install Kube-router addon.

What's next

Once you have installed the Kube-router addon, you can follow the Declare Network Policy to try out Kubernetes NetworkPolicy.

6.5 - Romana for NetworkPolicy

This page shows how to use Romana for NetworkPolicy.

Before you begin

Complete steps 1, 2, and 3 of the kubeadm getting started guide.

Installing Romana with kubeadm

Follow the containerized installation guide for kubeadm.

Applying network policies

To apply network policies use one of the following:

What's next

Once you have installed Romana, you can follow the Declare Network Policy to try out Kubernetes NetworkPolicy.

6.6 - Weave Net for NetworkPolicy

This page shows how to use Weave Net for NetworkPolicy.

Before you begin

You need to have a Kubernetes cluster. Follow the kubeadm getting started guide to bootstrap one.

Install the Weave Net addon

Follow the Integrating Kubernetes via the Addon guide.

The Weave Net addon for Kubernetes comes with a Network Policy Controller that automatically monitors Kubernetes for any NetworkPolicy annotations on all namespaces and configures iptables rules to allow or block traffic as directed by the policies.

Test the installation

Verify that the weave works.

Enter the following command:

kubectl get pods -n kube-system -o wide

The output is similar to this:

NAME                                    READY     STATUS    RESTARTS   AGE       IP              NODE
weave-net-1t1qg                         2/2       Running   0          9d        192.168.2.10    worknode3
weave-net-231d7                         2/2       Running   1          7d        10.2.0.17       worknodegpu
weave-net-7nmwt                         2/2       Running   3          9d        192.168.2.131   masternode
weave-net-pmw8w                         2/2       Running   0          9d        192.168.2.216   worknode2

Each Node has a weave Pod, and all Pods are Running and 2/2 READY. (2/2 means that each Pod has weave and weave-npc.)

What's next

Once you have installed the Weave Net addon, you can follow the Declare Network Policy to try out Kubernetes NetworkPolicy. If you have any question, contact us at #weave-community on Slack or Weave User Group.

7 - Access Clusters Using the Kubernetes API

This page shows how to access clusters using the Kubernetes API.

Before you begin

You need to have a Kubernetes cluster, and the kubectl command-line tool must be configured to communicate with your cluster. It is recommended to run this tutorial on a cluster with at least two nodes that are not acting as control plane hosts. If you do not already have a cluster, you can create one by using minikube or you can use one of these Kubernetes playgrounds:

To check the version, enter kubectl version.

Accessing the Kubernetes API

Accessing for the first time with kubectl

When accessing the Kubernetes API for the first time, use the Kubernetes command-line tool, kubectl.

To access a cluster, you need to know the location of the cluster and have credentials to access it. Typically, this is automatically set-up when you work through a Getting started guide, or someone else set up the cluster and provided you with credentials and a location.

Check the location and credentials that kubectl knows about with this command:

kubectl config view

Many of the examples provide an introduction to using kubectl. Complete documentation is found in the kubectl manual.

Directly accessing the REST API

kubectl handles locating and authenticating to the API server. If you want to directly access the REST API with an http client like curl or wget, or a browser, there are multiple ways you can locate and authenticate against the API server:

  1. Run kubectl in proxy mode (recommended). This method is recommended, since it uses the stored API server location and verifies the identity of the API server using a self-signed certificate. No man-in-the-middle (MITM) attack is possible using this method.
  2. Alternatively, you can provide the location and credentials directly to the http client. This works with client code that is confused by proxies. To protect against man in the middle attacks, you'll need to import a root cert into your browser.

Using the Go or Python client libraries provides accessing kubectl in proxy mode.

Using kubectl proxy

The following command runs kubectl in a mode where it acts as a reverse proxy. It handles locating the API server and authenticating.

Run it like this:

kubectl proxy --port=8080 &

See kubectl proxy for more details.

Then you can explore the API with curl, wget, or a browser, like so:

curl http://localhost:8080/api/

The output is similar to this:

{
  "versions": [
    "v1"
  ],
  "serverAddressByClientCIDRs": [
    {
      "clientCIDR": "0.0.0.0/0",
      "serverAddress": "10.0.1.149:443"
    }
  ]
}

Without kubectl proxy

It is possible to avoid using kubectl proxy by passing an authentication token directly to the API server, like this:

Using grep/cut approach:

# Check all possible clusters, as your .KUBECONFIG may have multiple contexts:
kubectl config view -o jsonpath='{"Cluster name\tServer\n"}{range .clusters[*]}{.name}{"\t"}{.cluster.server}{"\n"}{end}'

# Select name of cluster you want to interact with from above output:
export CLUSTER_NAME="some_server_name"

# Point to the API server referring the cluster name
APISERVER=$(kubectl config view -o jsonpath="{.clusters[?(@.name==\"$CLUSTER_NAME\")].cluster.server}")

# Create a secret to hold a token for the default service account
kubectl apply -f - <<EOF
apiVersion: v1
kind: Secret
metadata:
  name: default-token
  annotations:
    kubernetes.io/service-account.name: default
type: kubernetes.io/service-account-token
EOF

# Wait for the token controller to populate the secret with a token:
while ! kubectl describe secret default-token | grep -E '^token' >/dev/null; do
  echo "waiting for token..." >&2
  sleep 1
done

# Get the token value
TOKEN=$(kubectl get secret default-token -o jsonpath='{.data.token}' | base64 --decode)

# Explore the API with TOKEN
curl -X GET $APISERVER/api --header "Authorization: Bearer $TOKEN" --insecure

The output is similar to this:

{
  "kind": "APIVersions",
  "versions": [
    "v1"
  ],
  "serverAddressByClientCIDRs": [
    {
      "clientCIDR": "0.0.0.0/0",
      "serverAddress": "10.0.1.149:443"
    }
  ]
}

The above example uses the --insecure flag. This leaves it subject to MITM attacks. When kubectl accesses the cluster it uses a stored root certificate and client certificates to access the server. (These are installed in the ~/.kube directory). Since cluster certificates are typically self-signed, it may take special configuration to get your http client to use root certificate.

On some clusters, the API server does not require authentication; it may serve on localhost, or be protected by a firewall. There is not a standard for this. Controlling Access to the Kubernetes API describes how you can configure this as a cluster administrator.

Programmatic access to the API

Kubernetes officially supports client libraries for Go, Python, Java, dotnet, JavaScript, and Haskell. There are other client libraries that are provided and maintained by their authors, not the Kubernetes team. See client libraries for accessing the API from other languages and how they authenticate.

Go client

  • To get the library, run the following command: go get k8s.io/client-go@kubernetes-<kubernetes-version-number> See https://github.com/kubernetes/client-go/releases to see which versions are supported.
  • Write an application atop of the client-go clients.

The Go client can use the same kubeconfig file as the kubectl CLI does to locate and authenticate to the API server. See this example:

package main

import (
  "context"
  "fmt"
  "k8s.io/apimachinery/pkg/apis/meta/v1"
  "k8s.io/client-go/kubernetes"
  "k8s.io/client-go/tools/clientcmd"
)

func main() {
  // uses the current context in kubeconfig
  // path-to-kubeconfig -- for example, /root/.kube/config
  config, _ := clientcmd.BuildConfigFromFlags("", "<path-to-kubeconfig>")
  // creates the clientset
  clientset, _ := kubernetes.NewForConfig(config)
  // access the API to list pods
  pods, _ := clientset.CoreV1().Pods("").List(context.TODO(), v1.ListOptions{})
  fmt.Printf("There are %d pods in the cluster\n", len(pods.Items))
}

If the application is deployed as a Pod in the cluster, see Accessing the API from within a Pod.

Python client

To use Python client, run the following command: pip install kubernetes. See Python Client Library page for more installation options.

The Python client can use the same kubeconfig file as the kubectl CLI does to locate and authenticate to the API server. See this example:

from kubernetes import client, config

config.load_kube_config()

v1=client.CoreV1Api()
print("Listing pods with their IPs:")
ret = v1.list_pod_for_all_namespaces(watch=False)
for i in ret.items:
    print("%s\t%s\t%s" % (i.status.pod_ip, i.metadata.namespace, i.metadata.name))

Java client

To install the Java Client, run:

# Clone java library
git clone --recursive https://github.com/kubernetes-client/java

# Installing project artifacts, POM etc:
cd java
mvn install

See https://github.com/kubernetes-client/java/releases to see which versions are supported.

The Java client can use the same kubeconfig file as the kubectl CLI does to locate and authenticate to the API server. See this example:

package io.kubernetes.client.examples;

import io.kubernetes.client.ApiClient;
import io.kubernetes.client.ApiException;
import io.kubernetes.client.Configuration;
import io.kubernetes.client.apis.CoreV1Api;
import io.kubernetes.client.models.V1Pod;
import io.kubernetes.client.models.V1PodList;
import io.kubernetes.client.util.ClientBuilder;
import io.kubernetes.client.util.KubeConfig;
import java.io.FileReader;
import java.io.IOException;

/**
 * A simple example of how to use the Java API from an application outside a kubernetes cluster
 *
 * <p>Easiest way to run this: mvn exec:java
 * -Dexec.mainClass="io.kubernetes.client.examples.KubeConfigFileClientExample"
 *
 */
public class KubeConfigFileClientExample {
  public static void main(String[] args) throws IOException, ApiException {

    // file path to your KubeConfig
    String kubeConfigPath = "~/.kube/config";

    // loading the out-of-cluster config, a kubeconfig from file-system
    ApiClient client =
        ClientBuilder.kubeconfig(KubeConfig.loadKubeConfig(new FileReader(kubeConfigPath))).build();

    // set the global default api-client to the in-cluster one from above
    Configuration.setDefaultApiClient(client);

    // the CoreV1Api loads default api-client from global configuration.
    CoreV1Api api = new CoreV1Api();

    // invokes the CoreV1Api client
    V1PodList list = api.listPodForAllNamespaces(null, null, null, null, null, null, null, null, null);
    System.out.println("Listing all pods: ");
    for (V1Pod item : list.getItems()) {
      System.out.println(item.getMetadata().getName());
    }
  }
}

dotnet client

To use dotnet client, run the following command: dotnet add package KubernetesClient --version 1.6.1. See dotnet Client Library page for more installation options. See https://github.com/kubernetes-client/csharp/releases to see which versions are supported.

The dotnet client can use the same kubeconfig file as the kubectl CLI does to locate and authenticate to the API server. See this example:

using System;
using k8s;

namespace simple
{
    internal class PodList
    {
        private static void Main(string[] args)
        {
            var config = KubernetesClientConfiguration.BuildDefaultConfig();
            IKubernetes client = new Kubernetes(config);
            Console.WriteLine("Starting Request!");

            var list = client.ListNamespacedPod("default");
            foreach (var item in list.Items)
            {
                Console.WriteLine(item.Metadata.Name);
            }
            if (list.Items.Count == 0)
            {
                Console.WriteLine("Empty!");
            }
        }
    }
}

JavaScript client

To install JavaScript client, run the following command: npm install @kubernetes/client-node. See https://github.com/kubernetes-client/javascript/releases to see which versions are supported.

The JavaScript client can use the same kubeconfig file as the kubectl CLI does to locate and authenticate to the API server. See this example:

const k8s = require('@kubernetes/client-node');

const kc = new k8s.KubeConfig();
kc.loadFromDefault();

const k8sApi = kc.makeApiClient(k8s.CoreV1Api);

k8sApi.listNamespacedPod('default').then((res) => {
    console.log(res.body);
});

Haskell client

See https://github.com/kubernetes-client/haskell/releases to see which versions are supported.

The Haskell client can use the same kubeconfig file as the kubectl CLI does to locate and authenticate to the API server. See this example:

exampleWithKubeConfig :: IO ()
exampleWithKubeConfig = do
    oidcCache <- atomically $ newTVar $ Map.fromList []
    (mgr, kcfg) <- mkKubeClientConfig oidcCache $ KubeConfigFile "/path/to/kubeconfig"
    dispatchMime
            mgr
            kcfg
            (CoreV1.listPodForAllNamespaces (Accept MimeJSON))
        >>= print

What's next

8 - Advertise Extended Resources for a Node

This page shows how to specify extended resources for a Node. Extended resources allow cluster administrators to advertise node-level resources that would otherwise be unknown to Kubernetes.

Before you begin

You need to have a Kubernetes cluster, and the kubectl command-line tool must be configured to communicate with your cluster. It is recommended to run this tutorial on a cluster with at least two nodes that are not acting as control plane hosts. If you do not already have a cluster, you can create one by using minikube or you can use one of these Kubernetes playgrounds:

To check the version, enter kubectl version.

Get the names of your Nodes

kubectl get nodes

Choose one of your Nodes to use for this exercise.

To advertise a new extended resource on a Node, send an HTTP PATCH request to the Kubernetes API server. For example, suppose one of your Nodes has four dongles attached. Here's an example of a PATCH request that advertises four dongle resources for your Node.

PATCH /api/v1/nodes/<your-node-name>/status HTTP/1.1
Accept: application/json
Content-Type: application/json-patch+json
Host: k8s-master:8080

[
  {
    "op": "add",
    "path": "/status/capacity/example.com~1dongle",
    "value": "4"
  }
]

Note that Kubernetes does not need to know what a dongle is or what a dongle is for. The preceding PATCH request tells Kubernetes that your Node has four things that you call dongles.

Start a proxy, so that you can easily send requests to the Kubernetes API server:

kubectl proxy

In another command window, send the HTTP PATCH request. Replace <your-node-name> with the name of your Node:

curl --header "Content-Type: application/json-patch+json" \
  --request PATCH \
  --data '[{"op": "add", "path": "/status/capacity/example.com~1dongle", "value": "4"}]' \
  http://localhost:8001/api/v1/nodes/<your-node-name>/status

The output shows that the Node has a capacity of 4 dongles:

"capacity": {
  "cpu": "2",
  "memory": "2049008Ki",
  "example.com/dongle": "4",

Describe your Node:

kubectl describe node <your-node-name>

Once again, the output shows the dongle resource:

Capacity:
  cpu: 2
  memory: 2049008Ki
  example.com/dongle: 4

Now, application developers can create Pods that request a certain number of dongles. See Assign Extended Resources to a Container.

Discussion

Extended resources are similar to memory and CPU resources. For example, just as a Node has a certain amount of memory and CPU to be shared by all components running on the Node, it can have a certain number of dongles to be shared by all components running on the Node. And just as application developers can create Pods that request a certain amount of memory and CPU, they can create Pods that request a certain number of dongles.

Extended resources are opaque to Kubernetes; Kubernetes does not know anything about what they are. Kubernetes knows only that a Node has a certain number of them. Extended resources must be advertised in integer amounts. For example, a Node can advertise four dongles, but not 4.5 dongles.

Storage example

Suppose a Node has 800 GiB of a special kind of disk storage. You could create a name for the special storage, say example.com/special-storage. Then you could advertise it in chunks of a certain size, say 100 GiB. In that case, your Node would advertise that it has eight resources of type example.com/special-storage.

Capacity:
 ...
 example.com/special-storage: 8

If you want to allow arbitrary requests for special storage, you could advertise special storage in chunks of size 1 byte. In that case, you would advertise 800Gi resources of type example.com/special-storage.

Capacity:
 ...
 example.com/special-storage:  800Gi

Then a Container could request any number of bytes of special storage, up to 800Gi.

Clean up

Here is a PATCH request that removes the dongle advertisement from a Node.

PATCH /api/v1/nodes/<your-node-name>/status HTTP/1.1
Accept: application/json
Content-Type: application/json-patch+json
Host: k8s-master:8080

[
  {
    "op": "remove",
    "path": "/status/capacity/example.com~1dongle",
  }
]

Start a proxy, so that you can easily send requests to the Kubernetes API server:

kubectl proxy

In another command window, send the HTTP PATCH request. Replace <your-node-name> with the name of your Node:

curl --header "Content-Type: application/json-patch+json" \
  --request PATCH \
  --data '[{"op": "remove", "path": "/status/capacity/example.com~1dongle"}]' \
  http://localhost:8001/api/v1/nodes/<your-node-name>/status

Verify that the dongle advertisement has been removed:

kubectl describe node <your-node-name> | grep dongle

(you should not see any output)

What's next

For application developers

For cluster administrators

9 - Autoscale the DNS Service in a Cluster

This page shows how to enable and configure autoscaling of the DNS service in your Kubernetes cluster.

Before you begin

  • You need to have a Kubernetes cluster, and the kubectl command-line tool must be configured to communicate with your cluster. It is recommended to run this tutorial on a cluster with at least two nodes that are not acting as control plane hosts. If you do not already have a cluster, you can create one by using minikube or you can use one of these Kubernetes playgrounds:

    To check the version, enter kubectl version.

  • This guide assumes your nodes use the AMD64 or Intel 64 CPU architecture.

  • Make sure Kubernetes DNS is enabled.

Determine whether DNS horizontal autoscaling is already enabled

List the Deployments in your cluster in the kube-system namespace:

kubectl get deployment --namespace=kube-system

The output is similar to this:

NAME                   READY   UP-TO-DATE   AVAILABLE   AGE
...
kube-dns-autoscaler    1/1     1            1           ...
...

If you see "kube-dns-autoscaler" in the output, DNS horizontal autoscaling is already enabled, and you can skip to Tuning autoscaling parameters.

Get the name of your DNS Deployment

List the DNS deployments in your cluster in the kube-system namespace:

kubectl get deployment -l k8s-app=kube-dns --namespace=kube-system

The output is similar to this:

NAME      READY   UP-TO-DATE   AVAILABLE   AGE
...
coredns   2/2     2            2           ...
...

If you don't see a Deployment for DNS services, you can also look for it by name:

kubectl get deployment --namespace=kube-system

and look for a deployment named coredns or kube-dns.

Your scale target is

Deployment/<your-deployment-name>

where <your-deployment-name> is the name of your DNS Deployment. For example, if the name of your Deployment for DNS is coredns, your scale target is Deployment/coredns.

Enable DNS horizontal autoscaling

In this section, you create a new Deployment. The Pods in the Deployment run a container based on the cluster-proportional-autoscaler-amd64 image.

Create a file named dns-horizontal-autoscaler.yaml with this content:

admin/dns/dns-horizontal-autoscaler.yaml 
kind: ServiceAccount
apiVersion: v1
metadata:
  name: kube-dns-autoscaler
  namespace: kube-system
---
kind: ClusterRole
apiVersion: rbac.authorization.k8s.io/v1
metadata:
  name: system:kube-dns-autoscaler
rules:
  - apiGroups: [""]
    resources: ["nodes"]
    verbs: ["list", "watch"]
  - apiGroups: [""]
    resources: ["replicationcontrollers/scale"]
    verbs: ["get", "update"]
  - apiGroups: ["apps"]
    resources: ["deployments/scale", "replicasets/scale"]
    verbs: ["get", "update"]
# Remove the configmaps rule once below issue is fixed:
# kubernetes-incubator/cluster-proportional-autoscaler#16
  - apiGroups: [""]
    resources: ["configmaps"]
    verbs: ["get", "create"]
---
kind: ClusterRoleBinding
apiVersion: rbac.authorization.k8s.io/v1
metadata:
  name: system:kube-dns-autoscaler
subjects:
  - kind: ServiceAccount
    name: kube-dns-autoscaler
    namespace: kube-system
roleRef:
  kind: ClusterRole
  name: system:kube-dns-autoscaler
  apiGroup: rbac.authorization.k8s.io

---
apiVersion: apps/v1
kind: Deployment
metadata:
  name: kube-dns-autoscaler
  namespace: kube-system
  labels:
    k8s-app: kube-dns-autoscaler
    kubernetes.io/cluster-service: "true"
spec:
  selector:
    matchLabels:
      k8s-app: kube-dns-autoscaler
  template:
    metadata:
      labels:
        k8s-app: kube-dns-autoscaler
    spec:
      priorityClassName: system-cluster-critical
      securityContext:
        seccompProfile:
          type: RuntimeDefault
        supplementalGroups: [ 65534 ]
        fsGroup: 65534
      nodeSelector:
        kubernetes.io/os: linux
      containers:
      - name: autoscaler
        image: registry.k8s.io/cpa/cluster-proportional-autoscaler:1.8.4
        resources:
            requests:
                cpu: "20m"
                memory: "10Mi"
        command:
          - /cluster-proportional-autoscaler
          - --namespace=kube-system
          - --configmap=kube-dns-autoscaler
          # Should keep target in sync with cluster/addons/dns/kube-dns.yaml.base
          - --target=<SCALE_TARGET>
          # When cluster is using large nodes(with more cores), "coresPerReplica" should dominate.
          # If using small nodes, "nodesPerReplica" should dominate.
          - --default-params={"linear":{"coresPerReplica":256,"nodesPerReplica":16,"preventSinglePointFailure":true,"includeUnschedulableNodes":true}}
          - --logtostderr=true
          - --v=2
      tolerations:
      - key: "CriticalAddonsOnly"
        operator: "Exists"
      serviceAccountName: kube-dns-autoscaler

In the file, replace <SCALE_TARGET> with your scale target.

Go to the directory that contains your configuration file, and enter this command to create the Deployment:

kubectl apply -f dns-horizontal-autoscaler.yaml

The output of a successful command is:

deployment.apps/kube-dns-autoscaler created

DNS horizontal autoscaling is now enabled.

Tune DNS autoscaling parameters

Verify that the kube-dns-autoscaler ConfigMap exists:

kubectl get configmap --namespace=kube-system

The output is similar to this:

NAME                  DATA      AGE
...
kube-dns-autoscaler   1         ...
...

Modify the data in the ConfigMap:

kubectl edit configmap kube-dns-autoscaler --namespace=kube-system

Look for this line:

linear: '{"coresPerReplica":256,"min":1,"nodesPerReplica":16}'

Modify the fields according to your needs. The "min" field indicates the minimal number of DNS backends. The actual number of backends is calculated using this equation:

replicas = max( ceil( cores × 1/coresPerReplica ) , ceil( nodes × 1/nodesPerReplica ) )

Note that the values of both coresPerReplica and nodesPerReplica are floats.

The idea is that when a cluster is using nodes that have many cores, coresPerReplica dominates. When a cluster is using nodes that have fewer cores, nodesPerReplica dominates.

There are other supported scaling patterns. For details, see cluster-proportional-autoscaler.

Disable DNS horizontal autoscaling

There are a few options for tuning DNS horizontal autoscaling. Which option to use depends on different conditions.

Option 1: Scale down the kube-dns-autoscaler deployment to 0 replicas

This option works for all situations. Enter this command:

kubectl scale deployment --replicas=0 kube-dns-autoscaler --namespace=kube-system

The output is:

deployment.apps/kube-dns-autoscaler scaled

Verify that the replica count is zero:

kubectl get rs --namespace=kube-system

The output displays 0 in the DESIRED and CURRENT columns:

NAME                                  DESIRED   CURRENT   READY   AGE
...
kube-dns-autoscaler-6b59789fc8        0         0         0       ...
...

Option 2: Delete the kube-dns-autoscaler deployment

This option works if kube-dns-autoscaler is under your own control, which means no one will re-create it:

kubectl delete deployment kube-dns-autoscaler --namespace=kube-system

The output is:

deployment.apps "kube-dns-autoscaler" deleted

Option 3: Delete the kube-dns-autoscaler manifest file from the master node

This option works if kube-dns-autoscaler is under control of the (deprecated) Addon Manager, and you have write access to the master node.

Sign in to the master node and delete the corresponding manifest file. The common path for this kube-dns-autoscaler is:

/etc/kubernetes/addons/dns-horizontal-autoscaler/dns-horizontal-autoscaler.yaml

After the manifest file is deleted, the Addon Manager will delete the kube-dns-autoscaler Deployment.

Understanding how DNS horizontal autoscaling works

  • The cluster-proportional-autoscaler application is deployed separately from the DNS service.

  • An autoscaler Pod runs a client that polls the Kubernetes API server for the number of nodes and cores in the cluster.

  • A desired replica count is calculated and applied to the DNS backends based on the current schedulable nodes and cores and the given scaling parameters.

  • The scaling parameters and data points are provided via a ConfigMap to the autoscaler, and it refreshes its parameters table every poll interval to be up to date with the latest desired scaling parameters.

  • Changes to the scaling parameters are allowed without rebuilding or restarting the autoscaler Pod.

  • The autoscaler provides a controller interface to support two control patterns: linear and ladder.

What's next

10 - Change the Access Mode of a PersistentVolume to ReadWriteOncePod

This page shows how to change the access mode on an existing PersistentVolume to use ReadWriteOncePod.

Before you begin

You need to have a Kubernetes cluster, and the kubectl command-line tool must be configured to communicate with your cluster. It is recommended to run this tutorial on a cluster with at least two nodes that are not acting as control plane hosts. If you do not already have a cluster, you can create one by using minikube or you can use one of these Kubernetes playgrounds:

Your Kubernetes server must be at or later than version v1.22. To check the version, enter kubectl version.

Why should I use ReadWriteOncePod?

Prior to Kubernetes v1.22, the ReadWriteOnce access mode was commonly used to restrict PersistentVolume access for workloads that required single-writer access to storage. However, this access mode had a limitation: it restricted volume access to a single node, allowing multiple pods on the same node to read from and write to the same volume simultaneously. This could pose a risk for applications that demand strict single-writer access for data safety.

If ensuring single-writer access is critical for your workloads, consider migrating your volumes to ReadWriteOncePod.

Migrating existing PersistentVolumes

If you have existing PersistentVolumes, they can be migrated to use ReadWriteOncePod. Only migrations from ReadWriteOnce to ReadWriteOncePod are supported.

In this example, there is already a ReadWriteOnce "cat-pictures-pvc" PersistentVolumeClaim that is bound to a "cat-pictures-pv" PersistentVolume, and a "cat-pictures-writer" Deployment that uses this PersistentVolumeClaim.

And you can view the PVC before you make changes. Either view the manifest locally, or run kubectl get pvc <name-of-pvc> -o yaml. The output is similar to:

# cat-pictures-pvc.yaml
kind: PersistentVolumeClaim
apiVersion: v1
metadata:
  name: cat-pictures-pvc
spec:
  accessModes:
  - ReadWriteOnce
  resources:
    requests:
      storage: 1Gi

Here's an example Deployment that relies on that PersistentVolumeClaim:

# cat-pictures-writer-deployment.yaml
apiVersion: apps/v1
kind: Deployment
metadata:
  name: cat-pictures-writer
spec:
  replicas: 3
  selector:
    matchLabels:
      app: cat-pictures-writer
  template:
    metadata:
      labels:
        app: cat-pictures-writer
    spec:
      containers:
      - name: nginx
        image: nginx:1.14.2
        ports:
        - containerPort: 80
        volumeMounts:
        - name: cat-pictures
          mountPath: /mnt
      volumes:
      - name: cat-pictures
        persistentVolumeClaim:
          claimName: cat-pictures-pvc
          readOnly: false

As a first step, you need to edit your PersistentVolume's spec.persistentVolumeReclaimPolicy and set it to Retain. This ensures your PersistentVolume will not be deleted when you delete the corresponding PersistentVolumeClaim:

kubectl patch pv cat-pictures-pv -p '{"spec":{"persistentVolumeReclaimPolicy":"Retain"}}'

Next you need to stop any workloads that are using the PersistentVolumeClaim bound to the PersistentVolume you want to migrate, and then delete the PersistentVolumeClaim. Avoid making any other changes to the PersistentVolumeClaim, such as volume resizes, until after the migration is complete.

Once that is done, you need to clear your PersistentVolume's spec.claimRef.uid to ensure PersistentVolumeClaims can bind to it upon recreation:

kubectl scale --replicas=0 deployment cat-pictures-writer
kubectl delete pvc cat-pictures-pvc
kubectl patch pv cat-pictures-pv -p '{"spec":{"claimRef":{"uid":""}}}'

After that, replace the PersistentVolume's list of valid access modes to be (only) ReadWriteOncePod:

kubectl patch pv cat-pictures-pv -p '{"spec":{"accessModes":["ReadWriteOncePod"]}}'

Next you need to modify your PersistentVolumeClaim to set ReadWriteOncePod as the only access mode. You should also set the PersistentVolumeClaim's spec.volumeName to the name of your PersistentVolume to ensure it binds to this specific PersistentVolume.

Once this is done, you can recreate your PersistentVolumeClaim and start up your workloads:

# IMPORTANT: Make sure to edit your PVC in cat-pictures-pvc.yaml before applying. You need to:
# - Set ReadWriteOncePod as the only access mode
# - Set spec.volumeName to "cat-pictures-pv"

kubectl apply -f cat-pictures-pvc.yaml
kubectl apply -f cat-pictures-writer-deployment.yaml

Lastly you may edit your PersistentVolume's spec.persistentVolumeReclaimPolicy and set to it back to Delete if you previously changed it.

kubectl patch pv cat-pictures-pv -p '{"spec":{"persistentVolumeReclaimPolicy":"Delete"}}'

What's next

11 - Change the default StorageClass

This page shows how to change the default Storage Class that is used to provision volumes for PersistentVolumeClaims that have no special requirements.

Before you begin

You need to have a Kubernetes cluster, and the kubectl command-line tool must be configured to communicate with your cluster. It is recommended to run this tutorial on a cluster with at least two nodes that are not acting as control plane hosts. If you do not already have a cluster, you can create one by using minikube or you can use one of these Kubernetes playgrounds:

To check the version, enter kubectl version.

Why change the default storage class?

Depending on the installation method, your Kubernetes cluster may be deployed with an existing StorageClass that is marked as default. This default StorageClass is then used to dynamically provision storage for PersistentVolumeClaims that do not require any specific storage class. See PersistentVolumeClaim documentation for details.

The pre-installed default StorageClass may not fit well with your expected workload; for example, it might provision storage that is too expensive. If this is the case, you can either change the default StorageClass or disable it completely to avoid dynamic provisioning of storage.

Deleting the default StorageClass may not work, as it may be re-created automatically by the addon manager running in your cluster. Please consult the docs for your installation for details about addon manager and how to disable individual addons.

Changing the default StorageClass

  1. List the StorageClasses in your cluster:

    kubectl get storageclass

    The output is similar to this:

    NAME                 PROVISIONER               AGE
standard (default)   kubernetes.io/gce-pd      1d
gold                 kubernetes.io/gce-pd      1d

    The default StorageClass is marked by (default).

  2. Mark the default StorageClass as non-default:

    The default StorageClass has an annotation storageclass.kubernetes.io/is-default-class set to true. Any other value or absence of the annotation is interpreted as false.

    To mark a StorageClass as non-default, you need to change its value to false:

    kubectl patch storageclass standard -p '{"metadata": {"annotations":{"storageclass.kubernetes.io/is-default-class":"false"}}}'

    where standard is the name of your chosen StorageClass.

  3. Mark a StorageClass as default:

    Similar to the previous step, you need to add/set the annotation storageclass.kubernetes.io/is-default-class=true.

    kubectl patch storageclass gold -p '{"metadata": {"annotations":{"storageclass.kubernetes.io/is-default-class":"true"}}}'

    Please note that at most one StorageClass can be marked as default. If two or more of them are marked as default, a PersistentVolumeClaim without storageClassName explicitly specified cannot be created.

  4. Verify that your chosen StorageClass is default:

    kubectl get storageclass

    The output is similar to this:

    NAME             PROVISIONER               AGE
standard         kubernetes.io/gce-pd      1d
gold (default)   kubernetes.io/gce-pd      1d

What's next

12 - Switching from Polling to CRI Event-based Updates to Container Status

FEATURE STATE: Kubernetes v1.25 [alpha] (enabled by default: false)

This page shows how to migrate nodes to use event based updates for container status. The event-based implementation reduces node resource consumption by the kubelet, compared to the legacy approach that relies on polling. You may know this feature as evented Pod lifecycle event generator (PLEG). That's the name used internally within the Kubernetes project for a key implementation detail.

The polling based approach is referred to as generic PLEG.

Before you begin

  • You need to run a version of Kubernetes that provides this feature. Kubernetes v1.27 includes beta support for event-based container status updates. The feature is beta but is disabled by default because it requires support from the container runtime.
  • Your Kubernetes server must be at or later than version 1.26. To check the version, enter kubectl version. If you are running a different version of Kubernetes, check the documentation for that release.
  • The container runtime in use must support container lifecycle events. The kubelet automatically switches back to the legacy generic PLEG mechanism if the container runtime does not announce support for container lifecycle events, even if you have this feature gate enabled.

Why switch to Evented PLEG?

  • The Generic PLEG incurs non-negligible overhead due to frequent polling of container statuses.
  • This overhead is exacerbated by Kubelet's parallelized polling of container states, thus limiting its scalability and causing poor performance and reliability problems.
  • The goal of Evented PLEG is to reduce unnecessary work during inactivity by replacing periodic polling.

Switching to Evented PLEG

  1. Start the Kubelet with the feature gate EventedPLEG enabled. You can manage the kubelet feature gates editing the kubelet config file and restarting the kubelet service. You need to do this on each node where you are using this feature.

  2. Make sure the node is drained before proceeding.

  3. Start the container runtime with the container event generation enabled.

    Version 1.7+

    Version 1.26+

    Check if the CRI-O is already configured to emit CRI events by verifying the configuration,

    crio config | grep enable_pod_events

    If it is enabled, the output should be similar to the following:

    enable_pod_events = true

    To enable it, start the CRI-O daemon with the flag --enable-pod-events=true or use a dropin config with the following lines:

    [crio.runtime]
enable_pod_events: true
    Your Kubernetes server must be at or later than version 1.26. To check the version, enter kubectl version.

  4. Verify that the kubelet is using event-based container stage change monitoring. To check, look for the term EventedPLEG in the kubelet logs.

    The output should be similar to this:

    I0314 11:10:13.909915 1105457 feature_gate.go:249] feature gates: &{map[EventedPLEG:true]}

    If you have set --v to 4 and above, you might see more entries that indicate that the kubelet is using event-based container state monitoring.

    I0314 11:12:42.009542 1110177 evented.go:238] "Evented PLEG: Generated pod status from the received event" podUID=3b2c6172-b112-447a-ba96-94e7022912dc
I0314 11:12:44.623326 1110177 evented.go:238] "Evented PLEG: Generated pod status from the received event" podUID=b3fba5ea-a8c5-4b76-8f43-481e17e8ec40
I0314 11:12:44.714564 1110177 evented.go:238] "Evented PLEG: Generated pod status from the received event" podUID=b3fba5ea-a8c5-4b76-8f43-481e17e8ec40

What's next

13 - Change the Reclaim Policy of a PersistentVolume

This page shows how to change the reclaim policy of a Kubernetes PersistentVolume.

Before you begin

You need to have a Kubernetes cluster, and the kubectl command-line tool must be configured to communicate with your cluster. It is recommended to run this tutorial on a cluster with at least two nodes that are not acting as control plane hosts. If you do not already have a cluster, you can create one by using minikube or you can use one of these Kubernetes playgrounds:

To check the version, enter kubectl version.

Why change reclaim policy of a PersistentVolume

PersistentVolumes can have various reclaim policies, including "Retain", "Recycle", and "Delete". For dynamically provisioned PersistentVolumes, the default reclaim policy is "Delete". This means that a dynamically provisioned volume is automatically deleted when a user deletes the corresponding PersistentVolumeClaim. This automatic behavior might be inappropriate if the volume contains precious data. In that case, it is more appropriate to use the "Retain" policy. With the "Retain" policy, if a user deletes a PersistentVolumeClaim, the corresponding PersistentVolume will not be deleted. Instead, it is moved to the Released phase, where all of its data can be manually recovered.

Changing the reclaim policy of a PersistentVolume

  1. List the PersistentVolumes in your cluster:

    kubectl get pv

    The output is similar to this:

    NAME                                       CAPACITY   ACCESSMODES   RECLAIMPOLICY   STATUS    CLAIM             STORAGECLASS     REASON    AGE
pvc-b6efd8da-b7b5-11e6-9d58-0ed433a7dd94   4Gi        RWO           Delete          Bound     default/claim1    manual                     10s
pvc-b95650f8-b7b5-11e6-9d58-0ed433a7dd94   4Gi        RWO           Delete          Bound     default/claim2    manual                     6s
pvc-bb3ca71d-b7b5-11e6-9d58-0ed433a7dd94   4Gi        RWO           Delete          Bound     default/claim3    manual                     3s

    This list also includes the name of the claims that are bound to each volume for easier identification of dynamically provisioned volumes.

  2. Choose one of your PersistentVolumes and change its reclaim policy:

    kubectl patch pv <your-pv-name> -p '{"spec":{"persistentVolumeReclaimPolicy":"Retain"}}'

    where <your-pv-name> is the name of your chosen PersistentVolume.

  3. Verify that your chosen PersistentVolume has the right policy:

    kubectl get pv

    The output is similar to this:

    NAME                                       CAPACITY   ACCESSMODES   RECLAIMPOLICY   STATUS    CLAIM             STORAGECLASS     REASON    AGE
pvc-b6efd8da-b7b5-11e6-9d58-0ed433a7dd94   4Gi        RWO           Delete          Bound     default/claim1    manual                     40s
pvc-b95650f8-b7b5-11e6-9d58-0ed433a7dd94   4Gi        RWO           Delete          Bound     default/claim2    manual                     36s
pvc-bb3ca71d-b7b5-11e6-9d58-0ed433a7dd94   4Gi        RWO           Retain          Bound     default/claim3    manual                     33s

    In the preceding output, you can see that the volume bound to claim default/claim3 has reclaim policy Retain. It will not be automatically deleted when a user deletes claim default/claim3.

What's next

References

14 - Cloud Controller Manager Administration

FEATURE STATE: Kubernetes v1.11 [beta]

Since cloud providers develop and release at a different pace compared to the Kubernetes project, abstracting the provider-specific code to the cloud-controller-manager binary allows cloud vendors to evolve independently from the core Kubernetes code.

The cloud-controller-manager can be linked to any cloud provider that satisfies cloudprovider.Interface. For backwards compatibility, the cloud-controller-manager provided in the core Kubernetes project uses the same cloud libraries as kube-controller-manager. Cloud providers already supported in Kubernetes core are expected to use the in-tree cloud-controller-manager to transition out of Kubernetes core.

Administration

Requirements

Every cloud has their own set of requirements for running their own cloud provider integration, it should not be too different from the requirements when running kube-controller-manager. As a general rule of thumb you'll need:

  • cloud authentication/authorization: your cloud may require a token or IAM rules to allow access to their APIs
  • kubernetes authentication/authorization: cloud-controller-manager may need RBAC rules set to speak to the kubernetes apiserver
  • high availability: like kube-controller-manager, you may want a high available setup for cloud controller manager using leader election (on by default).

Running cloud-controller-manager

Successfully running cloud-controller-manager requires some changes to your cluster configuration.

  • kubelet, kube-apiserver, and kube-controller-manager must be set according to the user's usage of external CCM. If the user has an external CCM (not the internal cloud controller loops in the Kubernetes Controller Manager), then --cloud-provider=external must be specified. Otherwise, it should not be specified.

Keep in mind that setting up your cluster to use cloud controller manager will change your cluster behaviour in a few ways:

  • Components that specify --cloud-provider=external will add a taint node.cloudprovider.kubernetes.io/uninitialized with an effect NoSchedule during initialization. This marks the node as needing a second initialization from an external controller before it can be scheduled work. Note that in the event that cloud controller manager is not available, new nodes in the cluster will be left unschedulable. The taint is important since the scheduler may require cloud specific information about nodes such as their region or type (high cpu, gpu, high memory, spot instance, etc).
  • cloud information about nodes in the cluster will no longer be retrieved using local metadata, but instead all API calls to retrieve node information will go through cloud controller manager. This may mean you can restrict access to your cloud API on the kubelets for better security. For larger clusters you may want to consider if cloud controller manager will hit rate limits since it is now responsible for almost all API calls to your cloud from within the cluster.

The cloud controller manager can implement:

  • Node controller - responsible for updating kubernetes nodes using cloud APIs and deleting kubernetes nodes that were deleted on your cloud.
  • Service controller - responsible for loadbalancers on your cloud against services of type LoadBalancer.
  • Route controller - responsible for setting up network routes on your cloud
  • any other features you would like to implement if you are running an out-of-tree provider.

Examples

If you are using a cloud that is currently supported in Kubernetes core and would like to adopt cloud controller manager, see the cloud controller manager in kubernetes core.

For cloud controller managers not in Kubernetes core, you can find the respective projects in repositories maintained by cloud vendors or by SIGs.

For providers already in Kubernetes core, you can run the in-tree cloud controller manager as a DaemonSet in your cluster, use the following as a guideline:

admin/cloud/ccm-example.yaml 
# This is an example of how to set up cloud-controller-manager as a Daemonset in your cluster.
# It assumes that your masters can run pods and has the role node-role.kubernetes.io/master
# Note that this Daemonset will not work straight out of the box for your cloud, this is
# meant to be a guideline.

---
apiVersion: v1
kind: ServiceAccount
metadata:
  name: cloud-controller-manager
  namespace: kube-system
---
apiVersion: rbac.authorization.k8s.io/v1
kind: ClusterRoleBinding
metadata:
  name: system:cloud-controller-manager
roleRef:
  apiGroup: rbac.authorization.k8s.io
  kind: ClusterRole
  name: cluster-admin
subjects:
- kind: ServiceAccount
  name: cloud-controller-manager
  namespace: kube-system
---
apiVersion: apps/v1
kind: DaemonSet
metadata:
  labels:
    k8s-app: cloud-controller-manager
  name: cloud-controller-manager
  namespace: kube-system
spec:
  selector:
    matchLabels:
      k8s-app: cloud-controller-manager
  template:
    metadata:
      labels:
        k8s-app: cloud-controller-manager
    spec:
      serviceAccountName: cloud-controller-manager
      containers:
      - name: cloud-controller-manager
        # for in-tree providers we use registry.k8s.io/cloud-controller-manager
        # this can be replaced with any other image for out-of-tree providers
        image: registry.k8s.io/cloud-controller-manager:v1.8.0
        command:
        - /usr/local/bin/cloud-controller-manager
        - --cloud-provider=[YOUR_CLOUD_PROVIDER]  # Add your own cloud provider here!
        - --leader-elect=true
        - --use-service-account-credentials
        # these flags will vary for every cloud provider
        - --allocate-node-cidrs=true
        - --configure-cloud-routes=true
        - --cluster-cidr=172.17.0.0/16
      tolerations:
      # this is required so CCM can bootstrap itself
      - key: node.cloudprovider.kubernetes.io/uninitialized
        value: "true"
        effect: NoSchedule
      # these tolerations are to have the daemonset runnable on control plane nodes
      # remove them if your control plane nodes should not run pods
      - key: node-role.kubernetes.io/control-plane
        operator: Exists
        effect: NoSchedule
      - key: node-role.kubernetes.io/master
        operator: Exists
        effect: NoSchedule
      # this is to restrict CCM to only run on master nodes
      # the node selector may vary depending on your cluster setup
      nodeSelector:
        node-role.kubernetes.io/master: ""

Limitations

Running cloud controller manager comes with a few possible limitations. Although these limitations are being addressed in upcoming releases, it's important that you are aware of these limitations for production workloads.

Support for Volumes

Cloud controller manager does not implement any of the volume controllers found in kube-controller-manager as the volume integrations also require coordination with kubelets. As we evolve CSI (container storage interface) and add stronger support for flex volume plugins, necessary support will be added to cloud controller manager so that clouds can fully integrate with volumes. Learn more about out-of-tree CSI volume plugins here.

Scalability

The cloud-controller-manager queries your cloud provider's APIs to retrieve information for all nodes. For very large clusters, consider possible bottlenecks such as resource requirements and API rate limiting.

Chicken and Egg

The goal of the cloud controller manager project is to decouple development of cloud features from the core Kubernetes project. Unfortunately, many aspects of the Kubernetes project has assumptions that cloud provider features are tightly integrated into the project. As a result, adopting this new architecture can create several situations where a request is being made for information from a cloud provider, but the cloud controller manager may not be able to return that information without the original request being complete.

A good example of this is the TLS bootstrapping feature in the Kubelet. TLS bootstrapping assumes that the Kubelet has the ability to ask the cloud provider (or a local metadata service) for all its address types (private, public, etc) but cloud controller manager cannot set a node's address types without being initialized in the first place which requires that the kubelet has TLS certificates to communicate with the apiserver.

As this initiative evolves, changes will be made to address these issues in upcoming releases.

What's next

To build and develop your own cloud controller manager, read Developing Cloud Controller Manager.

15 - Configure a kubelet image credential provider

FEATURE STATE: Kubernetes v1.26 [stable]

Starting from Kubernetes v1.20, the kubelet can dynamically retrieve credentials for a container image registry using exec plugins. The kubelet and the exec plugin communicate through stdio (stdin, stdout, and stderr) using Kubernetes versioned APIs. These plugins allow the kubelet to request credentials for a container registry dynamically as opposed to storing static credentials on disk. For example, the plugin may talk to a local metadata server to retrieve short-lived credentials for an image that is being pulled by the kubelet.

You may be interested in using this capability if any of the below are true:

  • API calls to a cloud provider service are required to retrieve authentication information for a registry.
  • Credentials have short expiration times and requesting new credentials frequently is required.
  • Storing registry credentials on disk or in imagePullSecrets is not acceptable.

This guide demonstrates how to configure the kubelet's image credential provider plugin mechanism.

Before you begin

  • You need a Kubernetes cluster with nodes that support kubelet credential provider plugins. This support is available in Kubernetes 1.32; Kubernetes v1.24 and v1.25 included this as a beta feature, enabled by default.
  • A working implementation of a credential provider exec plugin. You can build your own plugin or use one provided by cloud providers.
Your Kubernetes server must be at or later than version v1.26. To check the version, enter kubectl version.

Installing Plugins on Nodes

A credential provider plugin is an executable binary that will be run by the kubelet. Ensure that the plugin binary exists on every node in your cluster and stored in a known directory. The directory will be required later when configuring kubelet flags.

Configuring the Kubelet

In order to use this feature, the kubelet expects two flags to be set:

  • --image-credential-provider-config - the path to the credential provider plugin config file.
  • --image-credential-provider-bin-dir - the path to the directory where credential provider plugin binaries are located.

Configure a kubelet credential provider

The configuration file passed into --image-credential-provider-config is read by the kubelet to determine which exec plugins should be invoked for which container images. Here's an example configuration file you may end up using if you are using the ECR-based plugin:

apiVersion: kubelet.config.k8s.io/v1
kind: CredentialProviderConfig
# providers is a list of credential provider helper plugins that will be enabled by the kubelet.
# Multiple providers may match against a single image, in which case credentials
# from all providers will be returned to the kubelet. If multiple providers are called
# for a single image, the results are combined. If providers return overlapping
# auth keys, the value from the provider earlier in this list is used.
providers:
  # name is the required name of the credential provider. It must match the name of the
  # provider executable as seen by the kubelet. The executable must be in the kubelet's
  # bin directory (set by the --image-credential-provider-bin-dir flag).
  - name: ecr-credential-provider
    # matchImages is a required list of strings used to match against images in order to
    # determine if this provider should be invoked. If one of the strings matches the
    # requested image from the kubelet, the plugin will be invoked and given a chance
    # to provide credentials. Images are expected to contain the registry domain
    # and URL path.
    #
    # Each entry in matchImages is a pattern which can optionally contain a port and a path.
    # Globs can be used in the domain, but not in the port or the path. Globs are supported
    # as subdomains like '*.k8s.io' or 'k8s.*.io', and top-level-domains such as 'k8s.*'.
    # Matching partial subdomains like 'app*.k8s.io' is also supported. Each glob can only match
    # a single subdomain segment, so `*.io` does **not** match `*.k8s.io`.
    #
    # A match exists between an image and a matchImage when all of the below are true:
    # - Both contain the same number of domain parts and each part matches.
    # - The URL path of an matchImages must be a prefix of the target image URL path.
    # - If the matchImages contains a port, then the port must match in the image as well.
    #
    # Example values of matchImages:
    # - 123456789.dkr.ecr.us-east-1.amazonaws.com
    # - *.azurecr.io
    # - gcr.io
    # - *.*.registry.io
    # - registry.io:8080/path
    matchImages:
      - "*.dkr.ecr.*.amazonaws.com"
      - "*.dkr.ecr.*.amazonaws.com.cn"
      - "*.dkr.ecr-fips.*.amazonaws.com"
      - "*.dkr.ecr.us-iso-east-1.c2s.ic.gov"
      - "*.dkr.ecr.us-isob-east-1.sc2s.sgov.gov"
    # defaultCacheDuration is the default duration the plugin will cache credentials in-memory
    # if a cache duration is not provided in the plugin response. This field is required.
    defaultCacheDuration: "12h"
    # Required input version of the exec CredentialProviderRequest. The returned CredentialProviderResponse
    # MUST use the same encoding version as the input. Current supported values are:
    # - credentialprovider.kubelet.k8s.io/v1
    apiVersion: credentialprovider.kubelet.k8s.io/v1
    # Arguments to pass to the command when executing it.
    # +optional
    # args:
    #   - --example-argument
    # Env defines additional environment variables to expose to the process. These
    # are unioned with the host's environment, as well as variables client-go uses
    # to pass argument to the plugin.
    # +optional
    env:
      - name: AWS_PROFILE
        value: example_profile

The providers field is a list of enabled plugins used by the kubelet. Each entry has a few required fields:

  • name: the name of the plugin which MUST match the name of the executable binary that exists in the directory passed into --image-credential-provider-bin-dir.
  • matchImages: a list of strings used to match against images in order to determine if this provider should be invoked. More on this below.
  • defaultCacheDuration: the default duration the kubelet will cache credentials in-memory if a cache duration was not specified by the plugin.
  • apiVersion: the API version that the kubelet and the exec plugin will use when communicating.

Each credential provider can also be given optional args and environment variables as well. Consult the plugin implementors to determine what set of arguments and environment variables are required for a given plugin.

Configure image matching

The matchImages field for each credential provider is used by the kubelet to determine whether a plugin should be invoked for a given image that a Pod is using. Each entry in matchImages is an image pattern which can optionally contain a port and a path. Globs can be used in the domain, but not in the port or the path. Globs are supported as subdomains like *.k8s.io or k8s.*.io, and top-level domains such as k8s.*. Matching partial subdomains like app*.k8s.io is also supported. Each glob can only match a single subdomain segment, so *.io does NOT match *.k8s.io.

A match exists between an image name and a matchImage entry when all of the below are true:

  • Both contain the same number of domain parts and each part matches.
  • The URL path of match image must be a prefix of the target image URL path.
  • If the matchImages contains a port, then the port must match in the image as well.

Some example values of matchImages patterns are:

  • 123456789.dkr.ecr.us-east-1.amazonaws.com
  • *.azurecr.io
  • gcr.io
  • *.*.registry.io
  • foo.registry.io:8080/path

What's next

16 - Configure Quotas for API Objects

This page shows how to configure quotas for API objects, including PersistentVolumeClaims and Services. A quota restricts the number of objects, of a particular type, that can be created in a namespace. You specify quotas in a ResourceQuota object.

Before you begin

You need to have a Kubernetes cluster, and the kubectl command-line tool must be configured to communicate with your cluster. It is recommended to run this tutorial on a cluster with at least two nodes that are not acting as control plane hosts. If you do not already have a cluster, you can create one by using minikube or you can use one of these Kubernetes playgrounds:

To check the version, enter kubectl version.

Create a namespace

Create a namespace so that the resources you create in this exercise are isolated from the rest of your cluster.

kubectl create namespace quota-object-example

Create a ResourceQuota

Here is the configuration file for a ResourceQuota object:

admin/resource/quota-objects.yaml 
apiVersion: v1
kind: ResourceQuota
metadata:
  name: object-quota-demo
spec:
  hard:
    persistentvolumeclaims: "1"
    services.loadbalancers: "2"
    services.nodeports: "0"

Create the ResourceQuota:

kubectl apply -f https://k8s.io/examples/admin/resource/quota-objects.yaml --namespace=quota-object-example

View detailed information about the ResourceQuota:

kubectl get resourcequota object-quota-demo --namespace=quota-object-example --output=yaml

The output shows that in the quota-object-example namespace, there can be at most one PersistentVolumeClaim, at most two Services of type LoadBalancer, and no Services of type NodePort.

status:
  hard:
    persistentvolumeclaims: "1"
    services.loadbalancers: "2"
    services.nodeports: "0"
  used:
    persistentvolumeclaims: "0"
    services.loadbalancers: "0"
    services.nodeports: "0"

Create a PersistentVolumeClaim

Here is the configuration file for a PersistentVolumeClaim object:

admin/resource/quota-objects-pvc.yaml 
apiVersion: v1
kind: PersistentVolumeClaim
metadata:
  name: pvc-quota-demo
spec:
  storageClassName: manual
  accessModes:
    - ReadWriteOnce
  resources:
    requests:
      storage: 3Gi

Create the PersistentVolumeClaim:

kubectl apply -f https://k8s.io/examples/admin/resource/quota-objects-pvc.yaml --namespace=quota-object-example

Verify that the PersistentVolumeClaim was created:

kubectl get persistentvolumeclaims --namespace=quota-object-example

The output shows that the PersistentVolumeClaim exists and has status Pending:

NAME             STATUS
pvc-quota-demo   Pending

Attempt to create a second PersistentVolumeClaim

Here is the configuration file for a second PersistentVolumeClaim:

admin/resource/quota-objects-pvc-2.yaml 
apiVersion: v1
kind: PersistentVolumeClaim
metadata:
  name: pvc-quota-demo-2
spec:
  storageClassName: manual
  accessModes:
    - ReadWriteOnce
  resources:
    requests:
      storage: 4Gi

Attempt to create the second PersistentVolumeClaim:

kubectl apply -f https://k8s.io/examples/admin/resource/quota-objects-pvc-2.yaml --namespace=quota-object-example

The output shows that the second PersistentVolumeClaim was not created, because it would have exceeded the quota for the namespace.

persistentvolumeclaims "pvc-quota-demo-2" is forbidden:
exceeded quota: object-quota-demo, requested: persistentvolumeclaims=1,
used: persistentvolumeclaims=1, limited: persistentvolumeclaims=1

Notes

These are the strings used to identify API resources that can be constrained by quotas:

StringAPI Object
"pods"Pod
"services"Service
"replicationcontrollers"ReplicationController
"resourcequotas"ResourceQuota
"secrets"Secret
"configmaps"ConfigMap
"persistentvolumeclaims"PersistentVolumeClaim
"services.nodeports"Service of type NodePort
"services.loadbalancers"Service of type LoadBalancer

Clean up

Delete your namespace:

kubectl delete namespace quota-object-example

What's next

For cluster administrators

For app developers

17 - Control CPU Management Policies on the Node

FEATURE STATE: Kubernetes v1.26 [stable]

Kubernetes keeps many aspects of how pods execute on nodes abstracted from the user. This is by design.  However, some workloads require stronger guarantees in terms of latency and/or performance in order to operate acceptably. The kubelet provides methods to enable more complex workload placement policies while keeping the abstraction free from explicit placement directives.

For detailed information on resource management, please refer to the Resource Management for Pods and Containers documentation.

Before you begin

You need to have a Kubernetes cluster, and the kubectl command-line tool must be configured to communicate with your cluster. It is recommended to run this tutorial on a cluster with at least two nodes that are not acting as control plane hosts. If you do not already have a cluster, you can create one by using minikube or you can use one of these Kubernetes playgrounds:

Your Kubernetes server must be at or later than version v1.26. To check the version, enter kubectl version.

If you are running an older version of Kubernetes, please look at the documentation for the version you are actually running.

CPU Management Policies

By default, the kubelet uses CFS quota to enforce pod CPU limits.  When the node runs many CPU-bound pods, the workload can move to different CPU cores depending on whether the pod is throttled and which CPU cores are available at scheduling time. Many workloads are not sensitive to this migration and thus work fine without any intervention.

However, in workloads where CPU cache affinity and scheduling latency significantly affect workload performance, the kubelet allows alternative CPU management policies to determine some placement preferences on the node.

Windows Support

FEATURE STATE: Kubernetes v1.32 [alpha] (enabled by default: false)

CPU Manager support can be enabled on Windows by using the WindowsCPUAndMemoryAffinity feature gate and it requires support in the container runtime. Once the feature gate is enabled, follow the steps below to configure the CPU manager policy.

Configuration

The CPU Manager policy is set with the --cpu-manager-policy kubelet flag or the cpuManagerPolicy field in KubeletConfiguration. There are two supported policies:

  • none: the default policy.
  • static: allows pods with certain resource characteristics to be granted increased CPU affinity and exclusivity on the node.

The CPU manager periodically writes resource updates through the CRI in order to reconcile in-memory CPU assignments with cgroupfs. The reconcile frequency is set through a new Kubelet configuration value --cpu-manager-reconcile-period. If not specified, it defaults to the same duration as --node-status-update-frequency.

The behavior of the static policy can be fine-tuned using the --cpu-manager-policy-options flag. The flag takes a comma-separated list of key=value policy options. If you disable the CPUManagerPolicyOptions feature gate then you cannot fine-tune CPU manager policies. In that case, the CPU manager operates only using its default settings.

In addition to the top-level CPUManagerPolicyOptions feature gate, the policy options are split into two groups: alpha quality (hidden by default) and beta quality (visible by default). The groups are guarded respectively by the CPUManagerPolicyAlphaOptions and CPUManagerPolicyBetaOptions feature gates. Diverging from the Kubernetes standard, these feature gates guard groups of options, because it would have been too cumbersome to add a feature gate for each individual option.

Changing the CPU Manager Policy

Since the CPU manager policy can only be applied when kubelet spawns new pods, simply changing from "none" to "static" won't apply to existing pods. So in order to properly change the CPU manager policy on a node, perform the following steps:

  1. Drain the node.
  2. Stop kubelet.
  3. Remove the old CPU manager state file. The path to this file is /var/lib/kubelet/cpu_manager_state by default. This clears the state maintained by the CPUManager so that the cpu-sets set up by the new policy won’t conflict with it.
  4. Edit the kubelet configuration to change the CPU manager policy to the desired value.
  5. Start kubelet.

Repeat this process for every node that needs its CPU manager policy changed. Skipping this process will result in kubelet crashlooping with the following error:

could not restore state from checkpoint: configured policy "static" differs from state checkpoint policy "none", please drain this node and delete the CPU manager checkpoint file "/var/lib/kubelet/cpu_manager_state" before restarting Kubelet

None policy

The none policy explicitly enables the existing default CPU affinity scheme, providing no affinity beyond what the OS scheduler does automatically.  Limits on CPU usage for Guaranteed pods and Burstable pods are enforced using CFS quota.

Static policy

The static policy allows containers in Guaranteed pods with integer CPU requests access to exclusive CPUs on the node. This exclusivity is enforced using the cpuset cgroup controller.

This policy manages a shared pool of CPUs that initially contains all CPUs in the node. The amount of exclusively allocatable CPUs is equal to the total number of CPUs in the node minus any CPU reservations by the kubelet --kube-reserved or --system-reserved options. From 1.17, the CPU reservation list can be specified explicitly by kubelet --reserved-cpus option. The explicit CPU list specified by --reserved-cpus takes precedence over the CPU reservation specified by --kube-reserved and --system-reserved. CPUs reserved by these options are taken, in integer quantity, from the initial shared pool in ascending order by physical core ID.  This shared pool is the set of CPUs on which any containers in BestEffort and Burstable pods run. Containers in Guaranteed pods with fractional CPU requests also run on CPUs in the shared pool. Only containers that are both part of a Guaranteed pod and have integer CPU requests are assigned exclusive CPUs.

As Guaranteed pods whose containers fit the requirements for being statically assigned are scheduled to the node, CPUs are removed from the shared pool and placed in the cpuset for the container. CFS quota is not used to bound the CPU usage of these containers as their usage is bound by the scheduling domain itself. In others words, the number of CPUs in the container cpuset is equal to the integer CPU limit specified in the pod spec. This static assignment increases CPU affinity and decreases context switches due to throttling for the CPU-bound workload.

Consider the containers in the following pod specs:

spec:
  containers:
  - name: nginx
    image: nginx

The pod above runs in the BestEffort QoS class because no resource requests or limits are specified. It runs in the shared pool.

spec:
  containers:
  - name: nginx
    image: nginx
    resources:
      limits:
        memory: "200Mi"
      requests:
        memory: "100Mi"

The pod above runs in the Burstable QoS class because resource requests do not equal limits and the cpu quantity is not specified. It runs in the shared pool.

spec:
  containers:
  - name: nginx
    image: nginx
    resources:
      limits:
        memory: "200Mi"
        cpu: "2"
      requests:
        memory: "100Mi"
        cpu: "1"

The pod above runs in the Burstable QoS class because resource requests do not equal limits. It runs in the shared pool.

spec:
  containers:
  - name: nginx
    image: nginx
    resources:
      limits:
        memory: "200Mi"
        cpu: "2"
      requests:
        memory: "200Mi"
        cpu: "2"

The pod above runs in the Guaranteed QoS class because requests are equal to limits. And the container's resource limit for the CPU resource is an integer greater than or equal to one. The nginx container is granted 2 exclusive CPUs.

spec:
  containers:
  - name: nginx
    image: nginx
    resources:
      limits:
        memory: "200Mi"
        cpu: "1.5"
      requests:
        memory: "200Mi"
        cpu: "1.5"

The pod above runs in the Guaranteed QoS class because requests are equal to limits. But the container's resource limit for the CPU resource is a fraction. It runs in the shared pool.

spec:
  containers:
  - name: nginx
    image: nginx
    resources:
      limits:
        memory: "200Mi"
        cpu: "2"

The pod above runs in the Guaranteed QoS class because only limits are specified and requests are set equal to limits when not explicitly specified. And the container's resource limit for the CPU resource is an integer greater than or equal to one. The nginx container is granted 2 exclusive CPUs.

Static policy options

You can toggle groups of options on and off based upon their maturity level using the following feature gates:

  • CPUManagerPolicyBetaOptions default enabled. Disable to hide beta-level options.
  • CPUManagerPolicyAlphaOptions default disabled. Enable to show alpha-level options. You will still have to enable each option using the CPUManagerPolicyOptions kubelet option.

The following policy options exist for the static CPUManager policy:

  • full-pcpus-only (beta, visible by default) (1.22 or higher)
  • distribute-cpus-across-numa (alpha, hidden by default) (1.23 or higher)
  • align-by-socket (alpha, hidden by default) (1.25 or higher)
  • distribute-cpus-across-cores (alpha, hidden by default) (1.31 or higher)
  • strict-cpu-reservation (alpha, hidden by default) (1.32 or higher)

If the full-pcpus-only policy option is specified, the static policy will always allocate full physical cores. By default, without this option, the static policy allocates CPUs using a topology-aware best-fit allocation. On SMT enabled systems, the policy can allocate individual virtual cores, which correspond to hardware threads. This can lead to different containers sharing the same physical cores; this behaviour in turn contributes to the noisy neighbours problem. With the option enabled, the pod will be admitted by the kubelet only if the CPU request of all its containers can be fulfilled by allocating full physical cores. If the pod does not pass the admission, it will be put in Failed state with the message SMTAlignmentError.

If the distribute-cpus-across-numapolicy option is specified, the static policy will evenly distribute CPUs across NUMA nodes in cases where more than one NUMA node is required to satisfy the allocation. By default, the CPUManager will pack CPUs onto one NUMA node until it is filled, with any remaining CPUs simply spilling over to the next NUMA node. This can cause undesired bottlenecks in parallel code relying on barriers (and similar synchronization primitives), as this type of code tends to run only as fast as its slowest worker (which is slowed down by the fact that fewer CPUs are available on at least one NUMA node). By distributing CPUs evenly across NUMA nodes, application developers can more easily ensure that no single worker suffers from NUMA effects more than any other, improving the overall performance of these types of applications.

If the align-by-socket policy option is specified, CPUs will be considered aligned at the socket boundary when deciding how to allocate CPUs to a container. By default, the CPUManager aligns CPU allocations at the NUMA boundary, which could result in performance degradation if CPUs need to be pulled from more than one NUMA node to satisfy the allocation. Although it tries to ensure that all CPUs are allocated from the minimum number of NUMA nodes, there is no guarantee that those NUMA nodes will be on the same socket. By directing the CPUManager to explicitly align CPUs at the socket boundary rather than the NUMA boundary, we are able to avoid such issues. Note, this policy option is not compatible with TopologyManager single-numa-node policy and does not apply to hardware where the number of sockets is greater than number of NUMA nodes.

If the distribute-cpus-across-cores policy option is specified, the static policy will attempt to allocate virtual cores (hardware threads) across different physical cores. By default, the CPUManager tends to pack cpus onto as few physical cores as possible, which can lead to contention among cpus on the same physical core and result in performance bottlenecks. By enabling the distribute-cpus-across-cores policy, the static policy ensures that cpus are distributed across as many physical cores as possible, reducing the contention on the same physical core and thereby improving overall performance. However, it is important to note that this strategy might be less effective when the system is heavily loaded. Under such conditions, the benefit of reducing contention diminishes. Conversely, default behavior can help in reducing inter-core communication overhead, potentially providing better performance under high load conditions.

The full-pcpus-only option can be enabled by adding full-pcpus-only=true to the CPUManager policy options. Likewise, the distribute-cpus-across-numa option can be enabled by adding distribute-cpus-across-numa=true to the CPUManager policy options. When both are set, they are "additive" in the sense that CPUs will be distributed across NUMA nodes in chunks of full-pcpus rather than individual cores. The align-by-socket policy option can be enabled by adding align-by-socket=true to the CPUManager policy options. It is also additive to the full-pcpus-only and distribute-cpus-across-numa policy options.

The distribute-cpus-across-cores option can be enabled by adding distribute-cpus-across-cores=true to the CPUManager policy options. It cannot be used with full-pcpus-only or distribute-cpus-across-numa policy options together at this moment.

The reservedSystemCPUs parameter in KubeletConfiguration, or the deprecated kubelet command line option --reserved-cpus, defines an explicit CPU set for OS system daemons and kubernetes system daemons. More details of this parameter can be found on the Explicitly Reserved CPU List page.

By default this isolation is implemented only for guaranteed pods with integer CPU requests not for burstable and best-effort pods (and guaranteed pods with fractional CPU requests). Admission is only comparing the cpu requests against the allocatable cpus. Since the cpu limit is higher than the request, the default behaviour allows burstable and best-effort pods to use up the capacity of reservedSystemCPUs and cause host OS services to starve in real life deployments.

If the strict-cpu-reservation policy option is enabled, the static policy will not allow any workload to use the CPU cores specified in reservedSystemCPUs.

The strict-cpu-reservation option can be enabled by adding strict-cpu-reservation=true to the CPUManager policy options followed by removing the /var/lib/kubelet/cpu_manager_state file and restart kubelet.

18 - Control Topology Management Policies on a node

FEATURE STATE: Kubernetes v1.27 [stable]

An increasing number of systems leverage a combination of CPUs and hardware accelerators to support latency-critical execution and high-throughput parallel computation. These include workloads in fields such as telecommunications, scientific computing, machine learning, financial services and data analytics. Such hybrid systems comprise a high performance environment.

In order to extract the best performance, optimizations related to CPU isolation, memory and device locality are required. However, in Kubernetes, these optimizations are handled by a disjoint set of components.

Topology Manager is a kubelet component that aims to coordinate the set of components that are responsible for these optimizations.

Before you begin

You need to have a Kubernetes cluster, and the kubectl command-line tool must be configured to communicate with your cluster. It is recommended to run this tutorial on a cluster with at least two nodes that are not acting as control plane hosts. If you do not already have a cluster, you can create one by using minikube or you can use one of these Kubernetes playgrounds:

Your Kubernetes server must be at or later than version v1.18. To check the version, enter kubectl version.

How topology manager works

Prior to the introduction of Topology Manager, the CPU and Device Manager in Kubernetes make resource allocation decisions independently of each other. This can result in undesirable allocations on multiple-socketed systems, and performance/latency sensitive applications will suffer due to these undesirable allocations. Undesirable in this case meaning, for example, CPUs and devices being allocated from different NUMA Nodes, thus incurring additional latency.

The Topology Manager is a kubelet component, which acts as a source of truth so that other kubelet components can make topology aligned resource allocation choices.

The Topology Manager provides an interface for components, called Hint Providers, to send and receive topology information. The Topology Manager has a set of node level policies which are explained below.

The Topology Manager receives topology information from the Hint Providers as a bitmask denoting NUMA Nodes available and a preferred allocation indication. The Topology Manager policies perform a set of operations on the hints provided and converge on the hint determined by the policy to give the optimal result. If an undesirable hint is stored, the preferred field for the hint will be set to false. In the current policies preferred is the narrowest preferred mask. The selected hint is stored as part of the Topology Manager. Depending on the policy configured, the pod can be accepted or rejected from the node based on the selected hint. The hint is then stored in the Topology Manager for use by the Hint Providers when making the resource allocation decisions.

Windows Support

FEATURE STATE: Kubernetes v1.32 [alpha] (enabled by default: false)

The Topology Manager support can be enabled on Windows by using the WindowsCPUAndMemoryAffinity feature gate and it requires support in the container runtime.

Topology manager scopes and policies

The Topology Manager currently:

  • aligns Pods of all QoS classes.
  • aligns the requested resources that Hint Provider provides topology hints for.

If these conditions are met, the Topology Manager will align the requested resources.

In order to customize how this alignment is carried out, the Topology Manager provides two distinct options: scope and policy.

The scope defines the granularity at which you would like resource alignment to be performed, for example, at the pod or container level. And the policy defines the actual policy used to carry out the alignment, for example, best-effort, restricted, and single-numa-node. Details on the various scopes and policies available today can be found below.

Topology manager scopes

The Topology Manager can deal with the alignment of resources in a couple of distinct scopes:

  • container (default)
  • pod

Either option can be selected at a time of the kubelet startup, by setting the topologyManagerScope in the kubelet configuration file.

container scope

The container scope is used by default. You can also explicitly set the topologyManagerScope to container in the kubelet configuration file.

Within this scope, the Topology Manager performs a number of sequential resource alignments, i.e., for each container (in a pod) a separate alignment is computed. In other words, there is no notion of grouping the containers to a specific set of NUMA nodes, for this particular scope. In effect, the Topology Manager performs an arbitrary alignment of individual containers to NUMA nodes.

The notion of grouping the containers was endorsed and implemented on purpose in the following scope, for example the pod scope.

pod scope

To select the pod scope, set topologyManagerScope in the kubelet configuration file to pod.

This scope allows for grouping all containers in a pod to a common set of NUMA nodes. That is, the Topology Manager treats a pod as a whole and attempts to allocate the entire pod (all containers) to either a single NUMA node or a common set of NUMA nodes. The following examples illustrate the alignments produced by the Topology Manager on different occasions:

  • all containers can be and are allocated to a single NUMA node;
  • all containers can be and are allocated to a shared set of NUMA nodes.

The total amount of particular resource demanded for the entire pod is calculated according to effective requests/limits formula, and thus, this total value is equal to the maximum of:

  • the sum of all app container requests,
  • the maximum of init container requests,

for a resource.

Using the pod scope in tandem with single-numa-node Topology Manager policy is specifically valuable for workloads that are latency sensitive or for high-throughput applications that perform IPC. By combining both options, you are able to place all containers in a pod onto a single NUMA node; hence, the inter-NUMA communication overhead can be eliminated for that pod.

In the case of single-numa-node policy, a pod is accepted only if a suitable set of NUMA nodes is present among possible allocations. Reconsider the example above:

  • a set containing only a single NUMA node - it leads to pod being admitted,
  • whereas a set containing more NUMA nodes - it results in pod rejection (because instead of one NUMA node, two or more NUMA nodes are required to satisfy the allocation).

To recap, the Topology Manager first computes a set of NUMA nodes and then tests it against the Topology Manager policy, which either leads to the rejection or admission of the pod.

Topology manager policies

The Topology Manager supports four allocation policies. You can set a policy via a kubelet flag, --topology-manager-policy. There are four supported policies:

  • none (default)
  • best-effort
  • restricted
  • single-numa-node

none policy

This is the default policy and does not perform any topology alignment.

best-effort policy

For each container in a Pod, the kubelet, with best-effort topology management policy, calls each Hint Provider to discover their resource availability. Using this information, the Topology Manager stores the preferred NUMA Node affinity for that container. If the affinity is not preferred, the Topology Manager will store this and admit the pod to the node anyway.

The Hint Providers can then use this information when making the resource allocation decision.

restricted policy

For each container in a Pod, the kubelet, with restricted topology management policy, calls each Hint Provider to discover their resource availability. Using this information, the Topology Manager stores the preferred NUMA Node affinity for that container. If the affinity is not preferred, the Topology Manager will reject this pod from the node. This will result in a pod entering a Terminated state with a pod admission failure.

Once the pod is in a Terminated state, the Kubernetes scheduler will not attempt to reschedule the pod. It is recommended to use a ReplicaSet or Deployment to trigger a redeployment of the pod. An external control loop could be also implemented to trigger a redeployment of pods that have the Topology Affinity error.

If the pod is admitted, the Hint Providers can then use this information when making the resource allocation decision.

single-numa-node policy

For each container in a Pod, the kubelet, with single-numa-node topology management policy, calls each Hint Provider to discover their resource availability. Using this information, the Topology Manager determines if a single NUMA Node affinity is possible. If it is, Topology Manager will store this and the Hint Providers can then use this information when making the resource allocation decision. If, however, this is not possible then the Topology Manager will reject the pod from the node. This will result in a pod in a Terminated state with a pod admission failure.

Once the pod is in a Terminated state, the Kubernetes scheduler will not attempt to reschedule the pod. It is recommended to use a Deployment with replicas to trigger a redeployment of the Pod. An external control loop could be also implemented to trigger a redeployment of pods that have the Topology Affinity error.

Topology manager policy options

Support for the Topology Manager policy options requires TopologyManagerPolicyOptions feature gate to be enabled (it is enabled by default).

You can toggle groups of options on and off based upon their maturity level using the following feature gates:

  • TopologyManagerPolicyBetaOptions default enabled. Enable to show beta-level options.
  • TopologyManagerPolicyAlphaOptions default disabled. Enable to show alpha-level options.

You will still have to enable each option using the TopologyManagerPolicyOptions kubelet option.

prefer-closest-numa-nodes

The prefer-closest-numa-nodes option is GA since Kubernetes 1.32. In Kubernetes 1.32 this policy option is visible by default provided that the TopologyManagerPolicyOptions feature gate is enabled.

The Topology Manager is not aware by default of NUMA distances, and does not take them into account when making Pod admission decisions. This limitation surfaces in multi-socket, as well as single-socket multi NUMA systems, and can cause significant performance degradation in latency-critical execution and high-throughput applications if the Topology Manager decides to align resources on non-adjacent NUMA nodes.

If you specify the prefer-closest-numa-nodes policy option, the best-effort and restricted policies favor sets of NUMA nodes with shorter distance between them when making admission decisions.

You can enable this option by adding prefer-closest-numa-nodes=true to the Topology Manager policy options.

By default (without this option), the Topology Manager aligns resources on either a single NUMA node or, in the case where more than one NUMA node is required, using the minimum number of NUMA nodes.

max-allowable-numa-nodes (beta)

The max-allowable-numa-nodes option is beta since Kubernetes 1.31. In Kubernetes 1.32, this policy option is visible by default provided that the TopologyManagerPolicyOptions and TopologyManagerPolicyBetaOptions feature gates are enabled.

The time to admit a pod is tied to the number of NUMA nodes on the physical machine. By default, Kubernetes does not run a kubelet with the Topology Manager enabled, on any (Kubernetes) node where more than 8 NUMA nodes are detected.

You can enable this option by adding max-allowable-numa-nodes=true to the Topology Manager policy options.

Setting a value of max-allowable-numa-nodes does not (in and of itself) affect the latency of pod admission, but binding a Pod to a (Kubernetes) node with many NUMA does have an impact. Future, potential improvements to Kubernetes may improve Pod admission performance and the high latency that happens as the number of NUMA nodes increases.

Pod interactions with topology manager policies

Consider the containers in the following Pod manifest:

spec:
  containers:
  - name: nginx
    image: nginx

This pod runs in the BestEffort QoS class because no resource requests or limits are specified.

spec:
  containers:
  - name: nginx
    image: nginx
    resources:
      limits:
        memory: "200Mi"
      requests:
        memory: "100Mi"

This pod runs in the Burstable QoS class because requests are less than limits.

If the selected policy is anything other than none, the Topology Manager would consider these Pod specifications. The Topology Manager would consult the Hint Providers to get topology hints. In the case of the static, the CPU Manager policy would return default topology hint, because these Pods do not explicitly request CPU resources.

spec:
  containers:
  - name: nginx
    image: nginx
    resources:
      limits:
        memory: "200Mi"
        cpu: "2"
        example.com/device: "1"
      requests:
        memory: "200Mi"
        cpu: "2"
        example.com/device: "1"

This pod with integer CPU request runs in the Guaranteed QoS class because requests are equal to limits.

spec:
  containers:
  - name: nginx
    image: nginx
    resources:
      limits:
        memory: "200Mi"
        cpu: "300m"
        example.com/device: "1"
      requests:
        memory: "200Mi"
        cpu: "300m"
        example.com/device: "1"

This pod with sharing CPU request runs in the Guaranteed QoS class because requests are equal to limits.

spec:
  containers:
  - name: nginx
    image: nginx
    resources:
      limits:
        example.com/deviceA: "1"
        example.com/deviceB: "1"
      requests:
        example.com/deviceA: "1"
        example.com/deviceB: "1"

This pod runs in the BestEffort QoS class because there are no CPU and memory requests.

The Topology Manager would consider the above pods. The Topology Manager would consult the Hint Providers, which are CPU and Device Manager to get topology hints for the pods.

In the case of the Guaranteed pod with integer CPU request, the static CPU Manager policy would return topology hints relating to the exclusive CPU and the Device Manager would send back hints for the requested device.

In the case of the Guaranteed pod with sharing CPU request, the static CPU Manager policy would return default topology hint as there is no exclusive CPU request and the Device Manager would send back hints for the requested device.

In the above two cases of the Guaranteed pod, the none CPU Manager policy would return default topology hint.

In the case of the BestEffort pod, the static CPU Manager policy would send back the default topology hint as there is no CPU request and the Device Manager would send back the hints for each of the requested devices.

Using this information the Topology Manager calculates the optimal hint for the pod and stores this information, which will be used by the Hint Providers when they are making their resource assignments.

Known limitations

  1. The maximum number of NUMA nodes that Topology Manager allows is 8. With more than 8 NUMA nodes, there will be a state explosion when trying to enumerate the possible NUMA affinities and generating their hints. See max-allowable-numa-nodes (beta) for more options.

  2. The scheduler is not topology-aware, so it is possible to be scheduled on a node and then fail on the node due to the Topology Manager.

19 - Customizing DNS Service

This page explains how to configure your DNS Pod(s) and customize the DNS resolution process in your cluster.

Before you begin

You need to have a Kubernetes cluster, and the kubectl command-line tool must be configured to communicate with your cluster. It is recommended to run this tutorial on a cluster with at least two nodes that are not acting as control plane hosts. If you do not already have a cluster, you can create one by using minikube or you can use one of these Kubernetes playgrounds:

Your cluster must be running the CoreDNS add-on.

Your Kubernetes server must be at or later than version v1.12. To check the version, enter kubectl version.

Introduction

DNS is a built-in Kubernetes service launched automatically using the addon manager cluster add-on.

If you are running CoreDNS as a Deployment, it will typically be exposed as a Kubernetes Service with a static IP address. The kubelet passes DNS resolver information to each container with the --cluster-dns=<dns-service-ip> flag.

DNS names also need domains. You configure the local domain in the kubelet with the flag --cluster-domain=<default-local-domain>.

The DNS server supports forward lookups (A and AAAA records), port lookups (SRV records), reverse IP address lookups (PTR records), and more. For more information, see DNS for Services and Pods.

If a Pod's dnsPolicy is set to default, it inherits the name resolution configuration from the node that the Pod runs on. The Pod's DNS resolution should behave the same as the node. But see Known issues.

If you don't want this, or if you want a different DNS config for pods, you can use the kubelet's --resolv-conf flag. Set this flag to "" to prevent Pods from inheriting DNS. Set it to a valid file path to specify a file other than /etc/resolv.conf for DNS inheritance.

CoreDNS

CoreDNS is a general-purpose authoritative DNS server that can serve as cluster DNS, complying with the DNS specifications.

CoreDNS ConfigMap options

CoreDNS is a DNS server that is modular and pluggable, with plugins adding new functionalities. The CoreDNS server can be configured by maintaining a Corefile, which is the CoreDNS configuration file. As a cluster administrator, you can modify the ConfigMap for the CoreDNS Corefile to change how DNS service discovery behaves for that cluster.

In Kubernetes, CoreDNS is installed with the following default Corefile configuration:

apiVersion: v1
kind: ConfigMap
metadata:
  name: coredns
  namespace: kube-system
data:
  Corefile: |
    .:53 {
        errors
        health {
            lameduck 5s
        }
        ready
        kubernetes cluster.local in-addr.arpa ip6.arpa {
            pods insecure
            fallthrough in-addr.arpa ip6.arpa
            ttl 30
        }
        prometheus :9153
        forward . /etc/resolv.conf
        cache 30
        loop
        reload
        loadbalance
    }

The Corefile configuration includes the following plugins of CoreDNS:

  • errors: Errors are logged to stdout.
  • health: Health of CoreDNS is reported to http://localhost:8080/health. In this extended syntax lameduck will make the process unhealthy then wait for 5 seconds before the process is shut down.
  • ready: An HTTP endpoint on port 8181 will return 200 OK, when all plugins that are able to signal readiness have done so.
  • kubernetes: CoreDNS will reply to DNS queries based on IP of the Services and Pods. You can find more details about this plugin on the CoreDNS website.
    • ttl allows you to set a custom TTL for responses. The default is 5 seconds. The minimum TTL allowed is 0 seconds, and the maximum is capped at 3600 seconds. Setting TTL to 0 will prevent records from being cached.
    • The pods insecure option is provided for backward compatibility with kube-dns.
    • You can use the pods verified option, which returns an A record only if there exists a pod in the same namespace with a matching IP.
    • The pods disabled option can be used if you don't use pod records.
  • prometheus: Metrics of CoreDNS are available at http://localhost:9153/metrics in the Prometheus format (also known as OpenMetrics).
  • forward: Any queries that are not within the Kubernetes cluster domain are forwarded to predefined resolvers (/etc/resolv.conf).
  • cache: This enables a frontend cache.
  • loop: Detects simple forwarding loops and halts the CoreDNS process if a loop is found.
  • reload: Allows automatic reload of a changed Corefile. After you edit the ConfigMap configuration, allow two minutes for your changes to take effect.
  • loadbalance: This is a round-robin DNS loadbalancer that randomizes the order of A, AAAA, and MX records in the answer.

You can modify the default CoreDNS behavior by modifying the ConfigMap.

Configuration of Stub-domain and upstream nameserver using CoreDNS

CoreDNS has the ability to configure stub-domains and upstream nameservers using the forward plugin.

Example

If a cluster operator has a Consul domain server located at "10.150.0.1", and all Consul names have the suffix ".consul.local". To configure it in CoreDNS, the cluster administrator creates the following stanza in the CoreDNS ConfigMap.

consul.local:53 {
    errors
    cache 30
    forward . 10.150.0.1
}

To explicitly force all non-cluster DNS lookups to go through a specific nameserver at 172.16.0.1, point the forward to the nameserver instead of /etc/resolv.conf

forward .  172.16.0.1

The final ConfigMap along with the default Corefile configuration looks like:

apiVersion: v1
kind: ConfigMap
metadata:
  name: coredns
  namespace: kube-system
data:
  Corefile: |
    .:53 {
        errors
        health
        kubernetes cluster.local in-addr.arpa ip6.arpa {
           pods insecure
           fallthrough in-addr.arpa ip6.arpa
        }
        prometheus :9153
        forward . 172.16.0.1
        cache 30
        loop
        reload
        loadbalance
    }
    consul.local:53 {
        errors
        cache 30
        forward . 10.150.0.1
    }

What's next

20 - Debugging DNS Resolution

This page provides hints on diagnosing DNS problems.

Before you begin

You need to have a Kubernetes cluster, and the kubectl command-line tool must be configured to communicate with your cluster. It is recommended to run this tutorial on a cluster with at least two nodes that are not acting as control plane hosts. If you do not already have a cluster, you can create one by using minikube or you can use one of these Kubernetes playgrounds:


Your cluster must be configured to use the CoreDNS addon or its precursor, kube-dns.

Your Kubernetes server must be at or later than version v1.6. To check the version, enter kubectl version.

Create a simple Pod to use as a test environment

admin/dns/dnsutils.yaml 
apiVersion: v1
kind: Pod
metadata:
  name: dnsutils
  namespace: default
spec:
  containers:
  - name: dnsutils
    image: registry.k8s.io/e2e-test-images/agnhost:2.39
    imagePullPolicy: IfNotPresent
  restartPolicy: Always

Use that manifest to create a Pod:

kubectl apply -f https://k8s.io/examples/admin/dns/dnsutils.yaml

pod/dnsutils created

…and verify its status:

kubectl get pods dnsutils

NAME       READY     STATUS    RESTARTS   AGE
dnsutils   1/1       Running   0          <some-time>

Once that Pod is running, you can exec nslookup in that environment. If you see something like the following, DNS is working correctly.

kubectl exec -i -t dnsutils -- nslookup kubernetes.default

Server:    10.0.0.10
Address 1: 10.0.0.10

Name:      kubernetes.default
Address 1: 10.0.0.1

If the nslookup command fails, check the following:

Check the local DNS configuration first

Take a look inside the resolv.conf file. (See Customizing DNS Service and Known issues below for more information)

kubectl exec -ti dnsutils -- cat /etc/resolv.conf

Verify that the search path and name server are set up like the following (note that search path may vary for different cloud providers):

search default.svc.cluster.local svc.cluster.local cluster.local google.internal c.gce_project_id.internal
nameserver 10.0.0.10
options ndots:5

Errors such as the following indicate a problem with the CoreDNS (or kube-dns) add-on or with associated Services:

kubectl exec -i -t dnsutils -- nslookup kubernetes.default

Server:    10.0.0.10
Address 1: 10.0.0.10

nslookup: can't resolve 'kubernetes.default'

or

kubectl exec -i -t dnsutils -- nslookup kubernetes.default

Server:    10.0.0.10
Address 1: 10.0.0.10 kube-dns.kube-system.svc.cluster.local

nslookup: can't resolve 'kubernetes.default'

Check if the DNS pod is running

Use the kubectl get pods command to verify that the DNS pod is running.

kubectl get pods --namespace=kube-system -l k8s-app=kube-dns

NAME                       READY     STATUS    RESTARTS   AGE
...
coredns-7b96bf9f76-5hsxb   1/1       Running   0           1h
coredns-7b96bf9f76-mvmmt   1/1       Running   0           1h
...

If you see that no CoreDNS Pod is running or that the Pod has failed/completed, the DNS add-on may not be deployed by default in your current environment and you will have to deploy it manually.

Check for errors in the DNS pod

Use the kubectl logs command to see logs for the DNS containers.

For CoreDNS:

kubectl logs --namespace=kube-system -l k8s-app=kube-dns

Here is an example of a healthy CoreDNS log:

.:53
2018/08/15 14:37:17 [INFO] CoreDNS-1.2.2
2018/08/15 14:37:17 [INFO] linux/amd64, go1.10.3, 2e322f6
CoreDNS-1.2.2
linux/amd64, go1.10.3, 2e322f6
2018/08/15 14:37:17 [INFO] plugin/reload: Running configuration MD5 = 24e6c59e83ce706f07bcc82c31b1ea1c

See if there are any suspicious or unexpected messages in the logs.

Is DNS service up?

Verify that the DNS service is up by using the kubectl get service command.

kubectl get svc --namespace=kube-system

NAME         TYPE        CLUSTER-IP     EXTERNAL-IP   PORT(S)             AGE
...
kube-dns     ClusterIP   10.0.0.10      <none>        53/UDP,53/TCP        1h
...

If you have created the Service or in the case it should be created by default but it does not appear, see debugging Services for more information.

Are DNS endpoints exposed?

You can verify that DNS endpoints are exposed by using the kubectl get endpoints command.

kubectl get endpoints kube-dns --namespace=kube-system

NAME       ENDPOINTS                       AGE
kube-dns   10.180.3.17:53,10.180.3.17:53    1h

If you do not see the endpoints, see the endpoints section in the debugging Services documentation.

For additional Kubernetes DNS examples, see the cluster-dns examples in the Kubernetes GitHub repository.

Are DNS queries being received/processed?

You can verify if queries are being received by CoreDNS by adding the log plugin to the CoreDNS configuration (aka Corefile). The CoreDNS Corefile is held in a ConfigMap named coredns. To edit it, use the command:

kubectl -n kube-system edit configmap coredns

Then add log in the Corefile section per the example below:

apiVersion: v1
kind: ConfigMap
metadata:
  name: coredns
  namespace: kube-system
data:
  Corefile: |
    .:53 {
        log
        errors
        health
        kubernetes cluster.local in-addr.arpa ip6.arpa {
          pods insecure
          upstream
          fallthrough in-addr.arpa ip6.arpa
        }
        prometheus :9153
        forward . /etc/resolv.conf
        cache 30
        loop
        reload
        loadbalance
    }

After saving the changes, it may take up to minute or two for Kubernetes to propagate these changes to the CoreDNS pods.

Next, make some queries and view the logs per the sections above in this document. If CoreDNS pods are receiving the queries, you should see them in the logs.

Here is an example of a query in the log:

.:53
2018/08/15 14:37:15 [INFO] CoreDNS-1.2.0
2018/08/15 14:37:15 [INFO] linux/amd64, go1.10.3, 2e322f6
CoreDNS-1.2.0
linux/amd64, go1.10.3, 2e322f6
2018/09/07 15:29:04 [INFO] plugin/reload: Running configuration MD5 = 162475cdf272d8aa601e6fe67a6ad42f
2018/09/07 15:29:04 [INFO] Reloading complete
172.17.0.18:41675 - [07/Sep/2018:15:29:11 +0000] 59925 "A IN kubernetes.default.svc.cluster.local. udp 54 false 512" NOERROR qr,aa,rd,ra 106 0.000066649s

Does CoreDNS have sufficient permissions?

CoreDNS must be able to list service and endpoint related resources to properly resolve service names.

Sample error message:

2022-03-18T07:12:15.699431183Z [INFO] 10.96.144.227:52299 - 3686 "A IN serverproxy.contoso.net.cluster.local. udp 52 false 512" SERVFAIL qr,aa,rd 145 0.000091221s

First, get the current ClusterRole of system:coredns:

kubectl describe clusterrole system:coredns -n kube-system

Expected output:

PolicyRule:
  Resources                        Non-Resource URLs  Resource Names  Verbs
  ---------                        -----------------  --------------  -----
  endpoints                        []                 []              [list watch]
  namespaces                       []                 []              [list watch]
  pods                             []                 []              [list watch]
  services                         []                 []              [list watch]
  endpointslices.discovery.k8s.io  []                 []              [list watch]

If any permissions are missing, edit the ClusterRole to add them:

kubectl edit clusterrole system:coredns -n kube-system

Example insertion of EndpointSlices permissions:

...
- apiGroups:
  - discovery.k8s.io
  resources:
  - endpointslices
  verbs:
  - list
  - watch
...

Are you in the right namespace for the service?

DNS queries that don't specify a namespace are limited to the pod's namespace.

If the namespace of the pod and service differ, the DNS query must include the namespace of the service.

This query is limited to the pod's namespace:

kubectl exec -i -t dnsutils -- nslookup <service-name>

This query specifies the namespace:

kubectl exec -i -t dnsutils -- nslookup <service-name>.<namespace>

To learn more about name resolution, see DNS for Services and Pods.

Known issues

Some Linux distributions (e.g. Ubuntu) use a local DNS resolver by default (systemd-resolved). Systemd-resolved moves and replaces /etc/resolv.conf with a stub file that can cause a fatal forwarding loop when resolving names in upstream servers. This can be fixed manually by using kubelet's --resolv-conf flag to point to the correct resolv.conf (With systemd-resolved, this is /run/systemd/resolve/resolv.conf). kubeadm automatically detects systemd-resolved, and adjusts the kubelet flags accordingly.

Kubernetes installs do not configure the nodes' resolv.conf files to use the cluster DNS by default, because that process is inherently distribution-specific. This should probably be implemented eventually.

Linux's libc (a.k.a. glibc) has a limit for the DNS nameserver records to 3 by default and Kubernetes needs to consume 1 nameserver record. This means that if a local installation already uses 3 nameservers, some of those entries will be lost. To work around this limit, the node can run dnsmasq, which will provide more nameserver entries. You can also use kubelet's --resolv-conf flag.

If you are using Alpine version 3.17 or earlier as your base image, DNS may not work properly due to a design issue with Alpine. Until musl version 1.24 didn't include TCP fallback to the DNS stub resolver meaning any DNS call above 512 bytes would fail. Please upgrade your images to Alpine version 3.18 or above.

What's next

21 - Declare Network Policy

This document helps you get started using the Kubernetes NetworkPolicy API to declare network policies that govern how pods communicate with each other.

Before you begin

You need to have a Kubernetes cluster, and the kubectl command-line tool must be configured to communicate with your cluster. It is recommended to run this tutorial on a cluster with at least two nodes that are not acting as control plane hosts. If you do not already have a cluster, you can create one by using minikube or you can use one of these Kubernetes playgrounds:

Your Kubernetes server must be at or later than version v1.8. To check the version, enter kubectl version.

Make sure you've configured a network provider with network policy support. There are a number of network providers that support NetworkPolicy, including:

Create an nginx deployment and expose it via a service

To see how Kubernetes network policy works, start off by creating an nginx Deployment.

kubectl create deployment nginx --image=nginx

deployment.apps/nginx created

Expose the Deployment through a Service called nginx.

kubectl expose deployment nginx --port=80

service/nginx exposed

The above commands create a Deployment with an nginx Pod and expose the Deployment through a Service named nginx. The nginx Pod and Deployment are found in the default namespace.

kubectl get svc,pod

NAME                        CLUSTER-IP    EXTERNAL-IP   PORT(S)    AGE
service/kubernetes          10.100.0.1    <none>        443/TCP    46m
service/nginx               10.100.0.16   <none>        80/TCP     33s

NAME                        READY         STATUS        RESTARTS   AGE
pod/nginx-701339712-e0qfq   1/1           Running       0          35s

Test the service by accessing it from another Pod

You should be able to access the new nginx service from other Pods. To access the nginx Service from another Pod in the default namespace, start a busybox container:

kubectl run busybox --rm -ti --image=busybox:1.28 -- /bin/sh

In your shell, run the following command:

wget --spider --timeout=1 nginx

Connecting to nginx (10.100.0.16:80)
remote file exists

Limit access to the nginx service

To limit the access to the nginx service so that only Pods with the label access: true can query it, create a NetworkPolicy object as follows:

service/networking/nginx-policy.yaml 
apiVersion: networking.k8s.io/v1
kind: NetworkPolicy
metadata:
  name: access-nginx
spec:
  podSelector:
    matchLabels:
      app: nginx
  ingress:
  - from:
    - podSelector:
        matchLabels:
          access: "true"

The name of a NetworkPolicy object must be a valid DNS subdomain name.

Assign the policy to the service

Use kubectl to create a NetworkPolicy from the above nginx-policy.yaml file:

kubectl apply -f https://k8s.io/examples/service/networking/nginx-policy.yaml

networkpolicy.networking.k8s.io/access-nginx created

Test access to the service when access label is not defined

When you attempt to access the nginx Service from a Pod without the correct labels, the request times out:

kubectl run busybox --rm -ti --image=busybox:1.28 -- /bin/sh

In your shell, run the command:

wget --spider --timeout=1 nginx

Connecting to nginx (10.100.0.16:80)
wget: download timed out

Define access label and test again

You can create a Pod with the correct labels to see that the request is allowed:

kubectl run busybox --rm -ti --labels="access=true" --image=busybox:1.28 -- /bin/sh

In your shell, run the command:

wget --spider --timeout=1 nginx

Connecting to nginx (10.100.0.16:80)
remote file exists

22 - Developing Cloud Controller Manager

FEATURE STATE: Kubernetes v1.11 [beta]

The cloud-controller-manager is a Kubernetes control plane component that embeds cloud-specific control logic. The cloud controller manager lets you link your cluster into your cloud provider's API, and separates out the components that interact with that cloud platform from components that only interact with your cluster.

By decoupling the interoperability logic between Kubernetes and the underlying cloud infrastructure, the cloud-controller-manager component enables cloud providers to release features at a different pace compared to the main Kubernetes project.

Background

Since cloud providers develop and release at a different pace compared to the Kubernetes project, abstracting the provider-specific code to the cloud-controller-manager binary allows cloud vendors to evolve independently from the core Kubernetes code.

The Kubernetes project provides skeleton cloud-controller-manager code with Go interfaces to allow you (or your cloud provider) to plug in your own implementations. This means that a cloud provider can implement a cloud-controller-manager by importing packages from Kubernetes core; each cloudprovider will register their own code by calling cloudprovider.RegisterCloudProvider to update a global variable of available cloud providers.

Developing

Out of tree

To build an out-of-tree cloud-controller-manager for your cloud:

  1. Create a go package with an implementation that satisfies cloudprovider.Interface.
  2. Use main.go in cloud-controller-manager from Kubernetes core as a template for your main.go. As mentioned above, the only difference should be the cloud package that will be imported.
  3. Import your cloud package in main.go, ensure your package has an init block to run cloudprovider.RegisterCloudProvider.

Many cloud providers publish their controller manager code as open source. If you are creating a new cloud-controller-manager from scratch, you could take an existing out-of-tree cloud controller manager as your starting point.

In tree

For in-tree cloud providers, you can run the in-tree cloud controller manager as a DaemonSet in your cluster. See Cloud Controller Manager Administration for more details.

23 - Enable Or Disable A Kubernetes API

This page shows how to enable or disable an API version from your cluster's control plane.

Specific API versions can be turned on or off by passing --runtime-config=api/<version> as a command line argument to the API server. The values for this argument are a comma-separated list of API versions. Later values override earlier values.

The runtime-config command line argument also supports 2 special keys:

  • api/all, representing all known APIs
  • api/legacy, representing only legacy APIs. Legacy APIs are any APIs that have been explicitly deprecated.

For example, to turn off all API versions except v1, pass --runtime-config=api/all=false,api/v1=true to the kube-apiserver.

What's next

Read the full documentation for the kube-apiserver component.

24 - Encrypting Confidential Data at Rest

All of the APIs in Kubernetes that let you write persistent API resource data support at-rest encryption. For example, you can enable at-rest encryption for Secrets. This at-rest encryption is additional to any system-level encryption for the etcd cluster or for the filesystem(s) on hosts where you are running the kube-apiserver.

This page shows how to enable and configure encryption of API data at rest.

Before you begin

  • You need to have a Kubernetes cluster, and the kubectl command-line tool must be configured to communicate with your cluster. It is recommended to run this tutorial on a cluster with at least two nodes that are not acting as control plane hosts. If you do not already have a cluster, you can create one by using minikube or you can use one of these Kubernetes playgrounds:

  • This task assumes that you are running the Kubernetes API server as a static pod on each control plane node.

  • Your cluster's control plane must use etcd v3.x (major version 3, any minor version).

  • To encrypt a custom resource, your cluster must be running Kubernetes v1.26 or newer.

  • To use a wildcard to match resources, your cluster must be running Kubernetes v1.27 or newer.

To check the version, enter kubectl version.

Determine whether encryption at rest is already enabled

By default, the API server stores plain-text representations of resources into etcd, with no at-rest encryption.

The kube-apiserver process accepts an argument --encryption-provider-config that specifies a path to a configuration file. The contents of that file, if you specify one, control how Kubernetes API data is encrypted in etcd. If you are running the kube-apiserver without the --encryption-provider-config command line argument, you do not have encryption at rest enabled. If you are running the kube-apiserver with the --encryption-provider-config command line argument, and the file that it references specifies the identity provider as the first encryption provider in the list, then you do not have at-rest encryption enabled (the default identity provider does not provide any confidentiality protection.)

If you are running the kube-apiserver with the --encryption-provider-config command line argument, and the file that it references specifies a provider other than identity as the first encryption provider in the list, then you already have at-rest encryption enabled. However, that check does not tell you whether a previous migration to encrypted storage has succeeded. If you are not sure, see ensure all relevant data are encrypted.

Understanding the encryption at rest configuration

---
#
# CAUTION: this is an example configuration.
#          Do not use this for your own cluster!
#
apiVersion: apiserver.config.k8s.io/v1
kind: EncryptionConfiguration
resources:
  - resources:
      - secrets
      - configmaps
      - pandas.awesome.bears.example # a custom resource API
    providers:
      # This configuration does not provide data confidentiality. The first
      # configured provider is specifying the "identity" mechanism, which
      # stores resources as plain text.
      #
      - identity: {} # plain text, in other words NO encryption
      - aesgcm:
          keys:
            - name: key1
              secret: c2VjcmV0IGlzIHNlY3VyZQ==
            - name: key2
              secret: dGhpcyBpcyBwYXNzd29yZA==
      - aescbc:
          keys:
            - name: key1
              secret: c2VjcmV0IGlzIHNlY3VyZQ==
            - name: key2
              secret: dGhpcyBpcyBwYXNzd29yZA==
      - secretbox:
          keys:
            - name: key1
              secret: YWJjZGVmZ2hpamtsbW5vcHFyc3R1dnd4eXoxMjM0NTY=
  - resources:
      - events
    providers:
      - identity: {} # do not encrypt Events even though *.* is specified below
  - resources:
      - '*.apps' # wildcard match requires Kubernetes 1.27 or later
    providers:
      - aescbc:
          keys:
          - name: key2
            secret: c2VjcmV0IGlzIHNlY3VyZSwgb3IgaXMgaXQ/Cg==
  - resources:
      - '*.*' # wildcard match requires Kubernetes 1.27 or later
    providers:
      - aescbc:
          keys:
          - name: key3
            secret: c2VjcmV0IGlzIHNlY3VyZSwgSSB0aGluaw==

Each resources array item is a separate config and contains a complete configuration. The resources.resources field is an array of Kubernetes resource names (resource or resource.group) that should be encrypted like Secrets, ConfigMaps, or other resources.

If custom resources are added to EncryptionConfiguration and the cluster version is 1.26 or newer, any newly created custom resources mentioned in the EncryptionConfiguration will be encrypted. Any custom resources that existed in etcd prior to that version and configuration will be unencrypted until they are next written to storage. This is the same behavior as built-in resources. See the Ensure all secrets are encrypted section.

The providers array is an ordered list of the possible encryption providers to use for the APIs that you listed. Each provider supports multiple keys - the keys are tried in order for decryption, and if the provider is the first provider, the first key is used for encryption.

Only one provider type may be specified per entry (identity or aescbc may be provided, but not both in the same item). The first provider in the list is used to encrypt resources written into the storage. When reading resources from storage, each provider that matches the stored data attempts in order to decrypt the data. If no provider can read the stored data due to a mismatch in format or secret key, an error is returned which prevents clients from accessing that resource.

EncryptionConfiguration supports the use of wildcards to specify the resources that should be encrypted. Use '*.<group>' to encrypt all resources within a group (for eg '*.apps' in above example) or '*.*' to encrypt all resources. '*.' can be used to encrypt all resource in the core group. '*.*' will encrypt all resources, even custom resources that are added after API server start.

If you have a wildcard covering resources and want to opt out of at-rest encryption for a particular kind of resource, you achieve that by adding a separate resources array item with the name of the resource that you want to exempt, followed by a providers array item where you specify the identity provider. You add this item to the list so that it appears earlier than the configuration where you do specify encryption (a provider that is not identity).

For example, if '*.*' is enabled and you want to opt out of encryption for Events and ConfigMaps, add a new earlier item to the resources, followed by the providers array item with identity as the provider. The more specific entry must come before the wildcard entry.

The new item would look similar to:

  ...
  - resources:
      - configmaps. # specifically from the core API group,
                    # because of trailing "."
      - events
    providers:
      - identity: {}
  # and then other entries in resources

Ensure that the exemption is listed before the wildcard '*.*' item in the resources array to give it precedence.

For more detailed information about the EncryptionConfiguration struct, please refer to the encryption configuration API.

Available providers

Before you configure encryption-at-rest for data in your cluster's Kubernetes API, you need to select which provider(s) you will use.

The following table describes each available provider.

Providers for Kubernetes encryption at rest
Name Encryption Strength Speed Key length
identity None N/A N/A N/A
Resources written as-is without encryption. When set as the first provider, the resource will be decrypted as new values are written. Existing encrypted resources are not automatically overwritten with the plaintext data. The identity provider is the default if you do not specify otherwise.
aescbc AES-CBC with PKCS#7 padding Weak Fast 32-byte
Not recommended due to CBC's vulnerability to padding oracle attacks. Key material accessible from control plane host.
aesgcm AES-GCM with random nonce Must be rotated every 200,000 writes Fastest 16, 24, or 32-byte
Not recommended for use except when an automated key rotation scheme is implemented. Key material accessible from control plane host.
kms v1 (deprecated since Kubernetes v1.28) Uses envelope encryption scheme with DEK per resource. Strongest Slow (compared to kms version 2) 32-bytes
Data is encrypted by data encryption keys (DEKs) using AES-GCM; DEKs are encrypted by key encryption keys (KEKs) according to configuration in Key Management Service (KMS). Simple key rotation, with a new DEK generated for each encryption, and KEK rotation controlled by the user.
Read how to configure the KMS V1 provider.
kms v2 Uses envelope encryption scheme with DEK per API server. Strongest Fast 32-bytes
Data is encrypted by data encryption keys (DEKs) using AES-GCM; DEKs are encrypted by key encryption keys (KEKs) according to configuration in Key Management Service (KMS). Kubernetes generates a new DEK per encryption from a secret seed. The seed is rotated whenever the KEK is rotated.
A good choice if using a third party tool for key management. Available as stable from Kubernetes v1.29.
Read how to configure the KMS V2 provider.
secretbox XSalsa20 and Poly1305 Strong Faster 32-byte
Uses relatively new encryption technologies that may not be considered acceptable in environments that require high levels of review. Key material accessible from control plane host.

The identity provider is the default if you do not specify otherwise. The identity provider does not encrypt stored data and provides no additional confidentiality protection.

Key storage

Local key storage

Encrypting secret data with a locally managed key protects against an etcd compromise, but it fails to protect against a host compromise. Since the encryption keys are stored on the host in the EncryptionConfiguration YAML file, a skilled attacker can access that file and extract the encryption keys.

Managed (KMS) key storage

The KMS provider uses envelope encryption: Kubernetes encrypts resources using a data key, and then encrypts that data key using the managed encryption service. Kubernetes generates a unique data key for each resource. The API server stores an encrypted version of the data key in etcd alongside the ciphertext; when reading the resource, the API server calls the managed encryption service and provides both the ciphertext and the (encrypted) data key. Within the managed encryption service, the provider use a key encryption key to decipher the data key, deciphers the data key, and finally recovers the plain text. Communication between the control plane and the KMS requires in-transit protection, such as TLS.

Using envelope encryption creates dependence on the key encryption key, which is not stored in Kubernetes. In the KMS case, an attacker who intends to get unauthorised access to the plaintext values would need to compromise etcd and the third-party KMS provider.

Protection for encryption keys

You should take appropriate measures to protect the confidential information that allows decryption, whether that is a local encryption key, or an authentication token that allows the API server to call KMS.

Even when you rely on a provider to manage the use and lifecycle of the main encryption key (or keys), you are still responsible for making sure that access controls and other security measures for the managed encryption service are appropriate for your security needs.

Encrypt your data

Generate the encryption key

The following steps assume that you are not using KMS, and therefore the steps also assume that you need to generate an encryption key. If you already have an encryption key, skip to Write an encryption configuration file.

Start by generating a new encryption key, and then encode it using base64:

Generate a 32-byte random key and base64 encode it. You can use this command:

head -c 32 /dev/urandom | base64

You can use /dev/hwrng instead of /dev/urandom if you want to use your PC's built-in hardware entropy source. Not all Linux devices provide a hardware random generator.

Generate a 32-byte random key and base64 encode it. You can use this command:

head -c 32 /dev/urandom | base64

Generate a 32-byte random key and base64 encode it. You can use this command:

# Do not run this in a session where you have set a random number
# generator seed.
[Convert]::ToBase64String((1..32|%{[byte](Get-Random -Max 256)}))

Replicate the encryption key

Using a secure mechanism for file transfer, make a copy of that encryption key available to every other control plane host.

At a minimum, use encryption in transit - for example, secure shell (SSH). For more security, use asymmetric encryption between hosts, or change the approach you are using so that you're relying on KMS encryption.

Write an encryption configuration file

Create a new encryption configuration file. The contents should be similar to:

---
apiVersion: apiserver.config.k8s.io/v1
kind: EncryptionConfiguration
resources:
  - resources:
      - secrets
      - configmaps
      - pandas.awesome.bears.example
    providers:
      - aescbc:
          keys:
            - name: key1
              # See the following text for more details about the secret value
              secret: <BASE 64 ENCODED SECRET>
      - identity: {} # this fallback allows reading unencrypted secrets;
                     # for example, during initial migration

To create a new encryption key (that does not use KMS), see Generate the encryption key.

Use the new encryption configuration file

You will need to mount the new encryption config file to the kube-apiserver static pod. Here is an example on how to do that:

  1. Save the new encryption config file to /etc/kubernetes/enc/enc.yaml on the control-plane node.

  2. Edit the manifest for the kube-apiserver static pod: /etc/kubernetes/manifests/kube-apiserver.yaml so that it is similar to:

    ---
#
# This is a fragment of a manifest for a static Pod.
# Check whether this is correct for your cluster and for your API server.
#
apiVersion: v1
kind: Pod
metadata:
  annotations:
    kubeadm.kubernetes.io/kube-apiserver.advertise-address.endpoint: 10.20.30.40:443
  creationTimestamp: null
  labels:
    app.kubernetes.io/component: kube-apiserver
    tier: control-plane
  name: kube-apiserver
  namespace: kube-system
spec:
  containers:
  - command:
    - kube-apiserver
    ...
    - --encryption-provider-config=/etc/kubernetes/enc/enc.yaml  # add this line
    volumeMounts:
    ...
    - name: enc                           # add this line
      mountPath: /etc/kubernetes/enc      # add this line
      readOnly: true                      # add this line
    ...
  volumes:
  ...
  - name: enc                             # add this line
    hostPath:                             # add this line
      path: /etc/kubernetes/enc           # add this line
      type: DirectoryOrCreate             # add this line
  ...

  3. Restart your API server.

You now have encryption in place for one control plane host. A typical Kubernetes cluster has multiple control plane hosts, so there is more to do.

Reconfigure other control plane hosts

If you have multiple API servers in your cluster, you should deploy the changes in turn to each API server.

When you are planning to update the encryption configuration of your cluster, plan this so that the API servers in your control plane can always decrypt the stored data (even part way through rolling out the change).

Make sure that you use the same encryption configuration on each control plane host.

Verify that newly written data is encrypted

Data is encrypted when written to etcd. After restarting your kube-apiserver, any newly created or updated Secret (or other resource kinds configured in EncryptionConfiguration) should be encrypted when stored.

To check this, you can use the etcdctl command line program to retrieve the contents of your secret data.

This example shows how to check this for encrypting the Secret API.

  1. Create a new Secret called secret1 in the default namespace:

    kubectl create secret generic secret1 -n default --from-literal=mykey=mydata

  2. Using the etcdctl command line tool, read that Secret out of etcd:

    ETCDCTL_API=3 etcdctl get /registry/secrets/default/secret1 [...] | hexdump -C

    where [...] must be the additional arguments for connecting to the etcd server.

    For example:

    ETCDCTL_API=3 etcdctl \
   --cacert=/etc/kubernetes/pki/etcd/ca.crt   \
   --cert=/etc/kubernetes/pki/etcd/server.crt \
   --key=/etc/kubernetes/pki/etcd/server.key  \
   get /registry/secrets/default/secret1 | hexdump -C

    The output is similar to this (abbreviated):

    00000000  2f 72 65 67 69 73 74 72  79 2f 73 65 63 72 65 74  |/registry/secret|
00000010  73 2f 64 65 66 61 75 6c  74 2f 73 65 63 72 65 74  |s/default/secret|
00000020  31 0a 6b 38 73 3a 65 6e  63 3a 61 65 73 63 62 63  |1.k8s:enc:aescbc|
00000030  3a 76 31 3a 6b 65 79 31  3a c7 6c e7 d3 09 bc 06  |:v1:key1:.l.....|
00000040  25 51 91 e4 e0 6c e5 b1  4d 7a 8b 3d b9 c2 7c 6e  |%Q...l..Mz.=..|n|
00000050  b4 79 df 05 28 ae 0d 8e  5f 35 13 2c c0 18 99 3e  |.y..(..._5.,...>|
[...]
00000110  23 3a 0d fc 28 ca 48 2d  6b 2d 46 cc 72 0b 70 4c  |#:..(.H-k-F.r.pL|
00000120  a5 fc 35 43 12 4e 60 ef  bf 6f fe cf df 0b ad 1f  |..5C.N`..o......|
00000130  82 c4 88 53 02 da 3e 66  ff 0a                    |...S..>f..|
0000013a

  3. Verify the stored Secret is prefixed with k8s:enc:aescbc:v1: which indicates the aescbc provider has encrypted the resulting data. Confirm that the key name shown in etcd matches the key name specified in the EncryptionConfiguration mentioned above. In this example, you can see that the encryption key named key1 is used in etcd and in EncryptionConfiguration.

  4. Verify the Secret is correctly decrypted when retrieved via the API:

    kubectl get secret secret1 -n default -o yaml

    The output should contain mykey: bXlkYXRh, with contents of mydata encoded using base64; read decoding a Secret to learn how to completely decode the Secret.

Ensure all relevant data are encrypted

It's often not enough to make sure that new objects get encrypted: you also want that encryption to apply to the objects that are already stored.

For this example, you have configured your cluster so that Secrets are encrypted on write. Performing a replace operation for each Secret will encrypt that content at rest, where the objects are unchanged.

You can make this change across all Secrets in your cluster:

# Run this as an administrator that can read and write all Secrets
kubectl get secrets --all-namespaces -o json | kubectl replace -f -

The command above reads all Secrets and then updates them with the same data, in order to apply server side encryption.

Prevent plain text retrieval

If you want to make sure that the only access to a particular API kind is done using encryption, you can remove the API server's ability to read that API's backing data as plaintext.

Once all Secrets in your cluster are encrypted, you can remove the identity part of the encryption configuration. For example:

---
apiVersion: apiserver.config.k8s.io/v1
kind: EncryptionConfiguration
resources:
  - resources:
      - secrets
    providers:
      - aescbc:
          keys:
            - name: key1
              secret: <BASE 64 ENCODED SECRET>
      - identity: {} # REMOVE THIS LINE

…and then restart each API server in turn. This change prevents the API server from accessing a plain-text Secret, even by accident.

Rotate a decryption key

Changing an encryption key for Kubernetes without incurring downtime requires a multi-step operation, especially in the presence of a highly-available deployment where multiple kube-apiserver processes are running.

  1. Generate a new key and add it as the second key entry for the current provider on all control plane nodes.
  2. Restart all kube-apiserver processes, to ensure each server can decrypt any data that are encrypted with the new key.
  3. Make a secure backup of the new encryption key. If you lose all copies of this key you would need to delete all the resources were encrypted under the lost key, and workloads may not operate as expected during the time that at-rest encryption is broken.
  4. Make the new key the first entry in the keys array so that it is used for encryption-at-rest for new writes
  5. Restart all kube-apiserver processes to ensure each control plane host now encrypts using the new key
  6. As a privileged user, run kubectl get secrets --all-namespaces -o json | kubectl replace -f - to encrypt all existing Secrets with the new key
  7. After you have updated all existing Secrets to use the new key and have made a secure backup of the new key, remove the old decryption key from the configuration.

Decrypt all data

This example shows how to stop encrypting the Secret API at rest. If you are encrypting other API kinds, adjust the steps to match.

To disable encryption at rest, place the identity provider as the first entry in your encryption configuration file:

---
apiVersion: apiserver.config.k8s.io/v1
kind: EncryptionConfiguration
resources:
  - resources:
      - secrets
      # list any other resources here that you previously were
      # encrypting at rest
    providers:
      - identity: {} # add this line
      - aescbc:
          keys:
            - name: key1
              secret: <BASE 64 ENCODED SECRET> # keep this in place
                                               # make sure it comes after "identity"

Then run the following command to force decryption of all Secrets:

kubectl get secrets --all-namespaces -o json | kubectl replace -f -

Once you have replaced all existing encrypted resources with backing data that don't use encryption, you can remove the encryption settings from the kube-apiserver.

Configure automatic reloading

You can configure automatic reloading of encryption provider configuration. That setting determines whether the API server should load the file you specify for --encryption-provider-config only once at startup, or automatically whenever you change that file. Enabling this option allows you to change the keys for encryption at rest without restarting the API server.

To allow automatic reloading, configure the API server to run with: --encryption-provider-config-automatic-reload=true. When enabled, file changes are polled every minute to observe the modifications. The apiserver_encryption_config_controller_automatic_reload_last_timestamp_seconds metric identifies when the new config becomes effective. This allows encryption keys to be rotated without restarting the API server.

What's next

25 - Decrypt Confidential Data that is Already Encrypted at Rest

All of the APIs in Kubernetes that let you write persistent API resource data support at-rest encryption. For example, you can enable at-rest encryption for Secrets. This at-rest encryption is additional to any system-level encryption for the etcd cluster or for the filesystem(s) on hosts where you are running the kube-apiserver.

This page shows how to switch from encryption of API data at rest, so that API data are stored unencrypted. You might want to do this to improve performance; usually, though, if it was a good idea to encrypt some data, it's also a good idea to leave them encrypted.

Before you begin

  • You need to have a Kubernetes cluster, and the kubectl command-line tool must be configured to communicate with your cluster. It is recommended to run this tutorial on a cluster with at least two nodes that are not acting as control plane hosts. If you do not already have a cluster, you can create one by using minikube or you can use one of these Kubernetes playgrounds:

  • This task assumes that you are running the Kubernetes API server as a static pod on each control plane node.

  • Your cluster's control plane must use etcd v3.x (major version 3, any minor version).

  • To encrypt a custom resource, your cluster must be running Kubernetes v1.26 or newer.

  • You should have some API data that are already encrypted.

To check the version, enter kubectl version.

Determine whether encryption at rest is already enabled

By default, the API server uses an identity provider that stores plain-text representations of resources. The default identity provider does not provide any confidentiality protection.

The kube-apiserver process accepts an argument --encryption-provider-config that specifies a path to a configuration file. The contents of that file, if you specify one, control how Kubernetes API data is encrypted in etcd. If it is not specified, you do not have encryption at rest enabled.

The format of that configuration file is YAML, representing a configuration API kind named EncryptionConfiguration. You can see an example configuration in Encryption at rest configuration.

If --encryption-provider-config is set, check which resources (such as secrets) are configured for encryption, and what provider is used. Make sure that the preferred provider for that resource type is not identity; you only set identity (no encryption) as default when you want to disable encryption at rest. Verify that the first-listed provider for a resource is something other than identity, which means that any new information written to resources of that type will be encrypted as configured. If you do see identity as the first-listed provider for any resource, this means that those resources are being written out to etcd without encryption.

Decrypt all data

This example shows how to stop encrypting the Secret API at rest. If you are encrypting other API kinds, adjust the steps to match.

Locate the encryption configuration file

First, find the API server configuration files. On each control plane node, static Pod manifest for the kube-apiserver specifies a command line argument, --encryption-provider-config. You are likely to find that this file is mounted into the static Pod using a hostPath volume mount. Once you locate the volume you can find the file on the node filesystem and inspect it.

Configure the API server to decrypt objects

To disable encryption at rest, place the identity provider as the first entry in your encryption configuration file.

For example, if your existing EncryptionConfiguration file reads:

---
apiVersion: apiserver.config.k8s.io/v1
kind: EncryptionConfiguration
resources:
  - resources:
      - secrets
    providers:
      - aescbc:
          keys:
            # Do not use this (invalid) example key for encryption
            - name: example
              secret: 2KfZgdiq2K0g2YrYpyDYs9mF2LPZhQ==

then change it to:

---
apiVersion: apiserver.config.k8s.io/v1
kind: EncryptionConfiguration
resources:
  - resources:
      - secrets
    providers:
      - identity: {} # add this line
      - aescbc:
          keys:
            - name: example
              secret: 2KfZgdiq2K0g2YrYpyDYs9mF2LPZhQ==

and restart the kube-apiserver Pod on this node.

Reconfigure other control plane hosts

If you have multiple API servers in your cluster, you should deploy the changes in turn to each API server.

Make sure that you use the same encryption configuration on each control plane host.

Force decryption

Then run the following command to force decryption of all Secrets:

# If you are decrypting a different kind of object, change "secrets" to match.
kubectl get secrets --all-namespaces -o json | kubectl replace -f -

Once you have replaced all existing encrypted resources with backing data that don't use encryption, you can remove the encryption settings from the kube-apiserver.

The command line options to remove are:

  • --encryption-provider-config
  • --encryption-provider-config-automatic-reload

Restart the kube-apiserver Pod again to apply the new configuration.

Reconfigure other control plane hosts

If you have multiple API servers in your cluster, you should again deploy the changes in turn to each API server.

Make sure that you use the same encryption configuration on each control plane host.

What's next

26 - Guaranteed Scheduling For Critical Add-On Pods

Kubernetes core components such as the API server, scheduler, and controller-manager run on a control plane node. However, add-ons must run on a regular cluster node. Some of these add-ons are critical to a fully functional cluster, such as metrics-server, DNS, and UI. A cluster may stop working properly if a critical add-on is evicted (either manually or as a side effect of another operation like upgrade) and becomes pending (for example when the cluster is highly utilized and either there are other pending pods that schedule into the space vacated by the evicted critical add-on pod or the amount of resources available on the node changed for some other reason).

Note that marking a pod as critical is not meant to prevent evictions entirely; it only prevents the pod from becoming permanently unavailable. A static pod marked as critical can't be evicted. However, non-static pods marked as critical are always rescheduled.

Marking pod as critical

To mark a Pod as critical, set priorityClassName for that Pod to system-cluster-critical or system-node-critical. system-node-critical is the highest available priority, even higher than system-cluster-critical.

27 - IP Masquerade Agent User Guide

This page shows how to configure and enable the ip-masq-agent.

Before you begin

You need to have a Kubernetes cluster, and the kubectl command-line tool must be configured to communicate with your cluster. It is recommended to run this tutorial on a cluster with at least two nodes that are not acting as control plane hosts. If you do not already have a cluster, you can create one by using minikube or you can use one of these Kubernetes playgrounds:

To check the version, enter kubectl version.

IP Masquerade Agent User Guide

The ip-masq-agent configures iptables rules to hide a pod's IP address behind the cluster node's IP address. This is typically done when sending traffic to destinations outside the cluster's pod CIDR range.

Key Terms

  • NAT (Network Address Translation): Is a method of remapping one IP address to another by modifying either the source and/or destination address information in the IP header. Typically performed by a device doing IP routing.
  • Masquerading: A form of NAT that is typically used to perform a many to one address translation, where multiple source IP addresses are masked behind a single address, which is typically the device doing the IP routing. In Kubernetes this is the Node's IP address.
  • CIDR (Classless Inter-Domain Routing): Based on the variable-length subnet masking, allows specifying arbitrary-length prefixes. CIDR introduced a new method of representation for IP addresses, now commonly known as CIDR notation, in which an address or routing prefix is written with a suffix indicating the number of bits of the prefix, such as 192.168.2.0/24.
  • Link Local: A link-local address is a network address that is valid only for communications within the network segment or the broadcast domain that the host is connected to. Link-local addresses for IPv4 are defined in the address block 169.254.0.0/16 in CIDR notation.

The ip-masq-agent configures iptables rules to handle masquerading node/pod IP addresses when sending traffic to destinations outside the cluster node's IP and the Cluster IP range. This essentially hides pod IP addresses behind the cluster node's IP address. In some environments, traffic to "external" addresses must come from a known machine address. For example, in Google Cloud, any traffic to the internet must come from a VM's IP. When containers are used, as in Google Kubernetes Engine, the Pod IP will be rejected for egress. To avoid this, we must hide the Pod IP behind the VM's own IP address - generally known as "masquerade". By default, the agent is configured to treat the three private IP ranges specified by RFC 1918 as non-masquerade CIDR. These ranges are 10.0.0.0/8, 172.16.0.0/12, and 192.168.0.0/16. The agent will also treat link-local (169.254.0.0/16) as a non-masquerade CIDR by default. The agent is configured to reload its configuration from the location /etc/config/ip-masq-agent every 60 seconds, which is also configurable.

masq/non-masq example

The agent configuration file must be written in YAML or JSON syntax, and may contain three optional keys:

  • nonMasqueradeCIDRs: A list of strings in CIDR notation that specify the non-masquerade ranges.
  • masqLinkLocal: A Boolean (true/false) which indicates whether to masquerade traffic to the link local prefix 169.254.0.0/16. False by default.
  • resyncInterval: A time interval at which the agent attempts to reload config from disk. For example: '30s', where 's' means seconds, 'ms' means milliseconds.

Traffic to 10.0.0.0/8, 172.16.0.0/12 and 192.168.0.0/16 ranges will NOT be masqueraded. Any other traffic (assumed to be internet) will be masqueraded. An example of a local destination from a pod could be its Node's IP address as well as another node's address or one of the IP addresses in Cluster's IP range. Any other traffic will be masqueraded by default. The below entries show the default set of rules that are applied by the ip-masq-agent:

iptables -t nat -L IP-MASQ-AGENT

target     prot opt source               destination
RETURN     all  --  anywhere             169.254.0.0/16       /* ip-masq-agent: cluster-local traffic should not be subject to MASQUERADE */ ADDRTYPE match dst-type !LOCAL
RETURN     all  --  anywhere             10.0.0.0/8           /* ip-masq-agent: cluster-local traffic should not be subject to MASQUERADE */ ADDRTYPE match dst-type !LOCAL
RETURN     all  --  anywhere             172.16.0.0/12        /* ip-masq-agent: cluster-local traffic should not be subject to MASQUERADE */ ADDRTYPE match dst-type !LOCAL
RETURN     all  --  anywhere             192.168.0.0/16       /* ip-masq-agent: cluster-local traffic should not be subject to MASQUERADE */ ADDRTYPE match dst-type !LOCAL
MASQUERADE  all  --  anywhere             anywhere             /* ip-masq-agent: outbound traffic should be subject to MASQUERADE (this match must come after cluster-local CIDR matches) */ ADDRTYPE match dst-type !LOCAL

By default, in GCE/Google Kubernetes Engine, if network policy is enabled or you are using a cluster CIDR not in the 10.0.0.0/8 range, the ip-masq-agent will run in your cluster. If you are running in another environment, you can add the ip-masq-agent DaemonSet to your cluster.

Create an ip-masq-agent

To create an ip-masq-agent, run the following kubectl command:

kubectl apply -f https://raw.githubusercontent.com/kubernetes-sigs/ip-masq-agent/master/ip-masq-agent.yaml

You must also apply the appropriate node label to any nodes in your cluster that you want the agent to run on.

kubectl label nodes my-node node.kubernetes.io/masq-agent-ds-ready=true

More information can be found in the ip-masq-agent documentation here.

In most cases, the default set of rules should be sufficient; however, if this is not the case for your cluster, you can create and apply a ConfigMap to customize the IP ranges that are affected. For example, to allow only 10.0.0.0/8 to be considered by the ip-masq-agent, you can create the following ConfigMap in a file called "config".

Run the following command to add the configmap to your cluster:

kubectl create configmap ip-masq-agent --from-file=config --namespace=kube-system

This will update a file located at /etc/config/ip-masq-agent which is periodically checked every resyncInterval and applied to the cluster node. After the resync interval has expired, you should see the iptables rules reflect your changes:

iptables -t nat -L IP-MASQ-AGENT

Chain IP-MASQ-AGENT (1 references)
target     prot opt source               destination
RETURN     all  --  anywhere             169.254.0.0/16       /* ip-masq-agent: cluster-local traffic should not be subject to MASQUERADE */ ADDRTYPE match dst-type !LOCAL
RETURN     all  --  anywhere             10.0.0.0/8           /* ip-masq-agent: cluster-local
MASQUERADE  all  --  anywhere             anywhere             /* ip-masq-agent: outbound traffic should be subject to MASQUERADE (this match must come after cluster-local CIDR matches) */ ADDRTYPE match dst-type !LOCAL

By default, the link local range (169.254.0.0/16) is also handled by the ip-masq agent, which sets up the appropriate iptables rules. To have the ip-masq-agent ignore link local, you can set masqLinkLocal to true in the ConfigMap.

nonMasqueradeCIDRs:
  - 10.0.0.0/8
resyncInterval: 60s
masqLinkLocal: true

28 - Limit Storage Consumption

This example demonstrates how to limit the amount of storage consumed in a namespace.

The following resources are used in the demonstration: ResourceQuota, LimitRange, and PersistentVolumeClaim.

Before you begin

  • You need to have a Kubernetes cluster, and the kubectl command-line tool must be configured to communicate with your cluster. It is recommended to run this tutorial on a cluster with at least two nodes that are not acting as control plane hosts. If you do not already have a cluster, you can create one by using minikube or you can use one of these Kubernetes playgrounds:

    To check the version, enter kubectl version.

Scenario: Limiting Storage Consumption

The cluster-admin is operating a cluster on behalf of a user population and the admin wants to control how much storage a single namespace can consume in order to control cost.

The admin would like to limit:

  1. The number of persistent volume claims in a namespace
  2. The amount of storage each claim can request
  3. The amount of cumulative storage the namespace can have

LimitRange to limit requests for storage

Adding a LimitRange to a namespace enforces storage request sizes to a minimum and maximum. Storage is requested via PersistentVolumeClaim. The admission controller that enforces limit ranges will reject any PVC that is above or below the values set by the admin.

In this example, a PVC requesting 10Gi of storage would be rejected because it exceeds the 2Gi max.

apiVersion: v1
kind: LimitRange
metadata:
  name: storagelimits
spec:
  limits:
  - type: PersistentVolumeClaim
    max:
      storage: 2Gi
    min:
      storage: 1Gi

Minimum storage requests are used when the underlying storage provider requires certain minimums. For example, AWS EBS volumes have a 1Gi minimum requirement.

ResourceQuota to limit PVC count and cumulative storage capacity

Admins can limit the number of PVCs in a namespace as well as the cumulative capacity of those PVCs. New PVCs that exceed either maximum value will be rejected.

In this example, a 6th PVC in the namespace would be rejected because it exceeds the maximum count of 5. Alternatively, a 5Gi maximum quota when combined with the 2Gi max limit above, cannot have 3 PVCs where each has 2Gi. That would be 6Gi requested for a namespace capped at 5Gi.

apiVersion: v1
kind: ResourceQuota
metadata:
  name: storagequota
spec:
  hard:
    persistentvolumeclaims: "5"
    requests.storage: "5Gi"

Summary

A limit range can put a ceiling on how much storage is requested while a resource quota can effectively cap the storage consumed by a namespace through claim counts and cumulative storage capacity. The allows a cluster-admin to plan their cluster's storage budget without risk of any one project going over their allotment.

29 - Migrate Replicated Control Plane To Use Cloud Controller Manager

The cloud-controller-manager is a Kubernetes control plane component that embeds cloud-specific control logic. The cloud controller manager lets you link your cluster into your cloud provider's API, and separates out the components that interact with that cloud platform from components that only interact with your cluster.

By decoupling the interoperability logic between Kubernetes and the underlying cloud infrastructure, the cloud-controller-manager component enables cloud providers to release features at a different pace compared to the main Kubernetes project.

Background

As part of the cloud provider extraction effort, all cloud specific controllers must be moved out of the kube-controller-manager. All existing clusters that run cloud controllers in the kube-controller-manager must migrate to instead run the controllers in a cloud provider specific cloud-controller-manager.

Leader Migration provides a mechanism in which HA clusters can safely migrate "cloud specific" controllers between the kube-controller-manager and the cloud-controller-manager via a shared resource lock between the two components while upgrading the replicated control plane. For a single-node control plane, or if unavailability of controller managers can be tolerated during the upgrade, Leader Migration is not needed and this guide can be ignored.

Leader Migration can be enabled by setting --enable-leader-migration on kube-controller-manager or cloud-controller-manager. Leader Migration only applies during the upgrade and can be safely disabled or left enabled after the upgrade is complete.

This guide walks you through the manual process of upgrading the control plane from kube-controller-manager with built-in cloud provider to running both kube-controller-manager and cloud-controller-manager. If you use a tool to deploy and manage the cluster, please refer to the documentation of the tool and the cloud provider for specific instructions of the migration.

Before you begin

It is assumed that the control plane is running Kubernetes version N and to be upgraded to version N + 1. Although it is possible to migrate within the same version, ideally the migration should be performed as part of an upgrade so that changes of configuration can be aligned to each release. The exact versions of N and N + 1 depend on each cloud provider. For example, if a cloud provider builds a cloud-controller-manager to work with Kubernetes 1.24, then N can be 1.23 and N + 1 can be 1.24.

The control plane nodes should run kube-controller-manager with Leader Election enabled, which is the default. As of version N, an in-tree cloud provider must be set with --cloud-provider flag and cloud-controller-manager should not yet be deployed.

The out-of-tree cloud provider must have built a cloud-controller-manager with Leader Migration implementation. If the cloud provider imports k8s.io/cloud-provider and k8s.io/controller-manager of version v0.21.0 or later, Leader Migration will be available. However, for version before v0.22.0, Leader Migration is alpha and requires feature gate ControllerManagerLeaderMigration to be enabled in cloud-controller-manager.

This guide assumes that kubelet of each control plane node starts kube-controller-manager and cloud-controller-manager as static pods defined by their manifests. If the components run in a different setting, please adjust the steps accordingly.

For authorization, this guide assumes that the cluster uses RBAC. If another authorization mode grants permissions to kube-controller-manager and cloud-controller-manager components, please grant the needed access in a way that matches the mode.

Grant access to Migration Lease

The default permissions of the controller manager allow only accesses to their main Lease. In order for the migration to work, accesses to another Lease are required.

You can grant kube-controller-manager full access to the leases API by modifying the system::leader-locking-kube-controller-manager role. This task guide assumes that the name of the migration lease is cloud-provider-extraction-migration.

kubectl patch -n kube-system role 'system::leader-locking-kube-controller-manager' -p '{"rules": [ {"apiGroups":[ "coordination.k8s.io"], "resources": ["leases"], "resourceNames": ["cloud-provider-extraction-migration"], "verbs": ["create", "list", "get", "update"] } ]}' --type=merge`

Do the same to the system::leader-locking-cloud-controller-manager role.

kubectl patch -n kube-system role 'system::leader-locking-cloud-controller-manager' -p '{"rules": [ {"apiGroups":[ "coordination.k8s.io"], "resources": ["leases"], "resourceNames": ["cloud-provider-extraction-migration"], "verbs": ["create", "list", "get", "update"] } ]}' --type=merge`

Initial Leader Migration configuration

Leader Migration optionally takes a configuration file representing the state of controller-to-manager assignment. At this moment, with in-tree cloud provider, kube-controller-manager runs route, service, and cloud-node-lifecycle. The following example configuration shows the assignment.

Leader Migration can be enabled without a configuration. Please see Default Configuration for details.

kind: LeaderMigrationConfiguration
apiVersion: controllermanager.config.k8s.io/v1
leaderName: cloud-provider-extraction-migration
controllerLeaders:
  - name: route
    component: kube-controller-manager
  - name: service
    component: kube-controller-manager
  - name: cloud-node-lifecycle
    component: kube-controller-manager

Alternatively, because the controllers can run under either controller managers, setting component to * for both sides makes the configuration file consistent between both parties of the migration.

# wildcard version
kind: LeaderMigrationConfiguration
apiVersion: controllermanager.config.k8s.io/v1
leaderName: cloud-provider-extraction-migration
controllerLeaders:
  - name: route
    component: *
  - name: service
    component: *
  - name: cloud-node-lifecycle
    component: *

On each control plane node, save the content to /etc/leadermigration.conf, and update the manifest of kube-controller-manager so that the file is mounted inside the container at the same location. Also, update the same manifest to add the following arguments:

  • --enable-leader-migration to enable Leader Migration on the controller manager
  • --leader-migration-config=/etc/leadermigration.conf to set configuration file

Restart kube-controller-manager on each node. At this moment, kube-controller-manager has leader migration enabled and is ready for the migration.

Deploy Cloud Controller Manager

In version N + 1, the desired state of controller-to-manager assignment can be represented by a new configuration file, shown as follows. Please note component field of each controllerLeaders changing from kube-controller-manager to cloud-controller-manager. Alternatively, use the wildcard version mentioned above, which has the same effect.

kind: LeaderMigrationConfiguration
apiVersion: controllermanager.config.k8s.io/v1
leaderName: cloud-provider-extraction-migration
controllerLeaders:
  - name: route
    component: cloud-controller-manager
  - name: service
    component: cloud-controller-manager
  - name: cloud-node-lifecycle
    component: cloud-controller-manager

When creating control plane nodes of version N + 1, the content should be deployed to /etc/leadermigration.conf. The manifest of cloud-controller-manager should be updated to mount the configuration file in the same manner as kube-controller-manager of version N. Similarly, add --enable-leader-migration and --leader-migration-config=/etc/leadermigration.conf to the arguments of cloud-controller-manager.

Create a new control plane node of version N + 1 with the updated cloud-controller-manager manifest, and with the --cloud-provider flag set to external for kube-controller-manager. kube-controller-manager of version N + 1 MUST NOT have Leader Migration enabled because, with an external cloud provider, it does not run the migrated controllers anymore, and thus it is not involved in the migration.

Please refer to Cloud Controller Manager Administration for more detail on how to deploy cloud-controller-manager.

Upgrade Control Plane

The control plane now contains nodes of both version N and N + 1. The nodes of version N run kube-controller-manager only, and these of version N + 1 run both kube-controller-manager and cloud-controller-manager. The migrated controllers, as specified in the configuration, are running under either kube-controller-manager of version N or cloud-controller-manager of version N + 1 depending on which controller manager holds the migration lease. No controller will ever be running under both controller managers at any time.

In a rolling manner, create a new control plane node of version N + 1 and bring down one of version N until the control plane contains only nodes of version N + 1. If a rollback from version N + 1 to N is required, add nodes of version N with Leader Migration enabled for kube-controller-manager back to the control plane, replacing one of version N + 1 each time until there are only nodes of version N.

(Optional) Disable Leader Migration

Now that the control plane has been upgraded to run both kube-controller-manager and cloud-controller-manager of version N + 1, Leader Migration has finished its job and can be safely disabled to save one Lease resource. It is safe to re-enable Leader Migration for the rollback in the future.

In a rolling manager, update manifest of cloud-controller-manager to unset both --enable-leader-migration and --leader-migration-config= flag, also remove the mount of /etc/leadermigration.conf, and finally remove /etc/leadermigration.conf. To re-enable Leader Migration, recreate the configuration file and add its mount and the flags that enable Leader Migration back to cloud-controller-manager.

Default Configuration

Starting Kubernetes 1.22, Leader Migration provides a default configuration suitable for the default controller-to-manager assignment. The default configuration can be enabled by setting --enable-leader-migration but without --leader-migration-config=.

For kube-controller-manager and cloud-controller-manager, if there are no flags that enable any in-tree cloud provider or change ownership of controllers, the default configuration can be used to avoid manual creation of the configuration file.

Special case: migrating the Node IPAM controller

If your cloud provider provides an implementation of Node IPAM controller, you should switch to the implementation in cloud-controller-manager. Disable Node IPAM controller in kube-controller-manager of version N + 1 by adding --controllers=*,-nodeipam to its flags. Then add nodeipam to the list of migrated controllers.

# wildcard version, with nodeipam
kind: LeaderMigrationConfiguration
apiVersion: controllermanager.config.k8s.io/v1
leaderName: cloud-provider-extraction-migration
controllerLeaders:
  - name: route
    component: *
  - name: service
    component: *
  - name: cloud-node-lifecycle
    component: *
  - name: nodeipam
-   component: *

What's next

30 - Namespaces Walkthrough

Kubernetes namespaces help different projects, teams, or customers to share a Kubernetes cluster.

It does this by providing the following:

  1. A scope for Names.
  2. A mechanism to attach authorization and policy to a subsection of the cluster.

Use of multiple namespaces is optional.

This example demonstrates how to use Kubernetes namespaces to subdivide your cluster.

Before you begin

You need to have a Kubernetes cluster, and the kubectl command-line tool must be configured to communicate with your cluster. It is recommended to run this tutorial on a cluster with at least two nodes that are not acting as control plane hosts. If you do not already have a cluster, you can create one by using minikube or you can use one of these Kubernetes playgrounds:

To check the version, enter kubectl version.

Prerequisites

This example assumes the following:

  1. You have an existing Kubernetes cluster.
  2. You have a basic understanding of Kubernetes Pods, Services, and Deployments.

Understand the default namespace

By default, a Kubernetes cluster will instantiate a default namespace when provisioning the cluster to hold the default set of Pods, Services, and Deployments used by the cluster.

Assuming you have a fresh cluster, you can inspect the available namespaces by doing the following:

kubectl get namespaces

NAME      STATUS    AGE
default   Active    13m

Create new namespaces

For this exercise, we will create two additional Kubernetes namespaces to hold our content.

Let's imagine a scenario where an organization is using a shared Kubernetes cluster for development and production use cases.

The development team would like to maintain a space in the cluster where they can get a view on the list of Pods, Services, and Deployments they use to build and run their application. In this space, Kubernetes resources come and go, and the restrictions on who can or cannot modify resources are relaxed to enable agile development.

The operations team would like to maintain a space in the cluster where they can enforce strict procedures on who can or cannot manipulate the set of Pods, Services, and Deployments that run the production site.

One pattern this organization could follow is to partition the Kubernetes cluster into two namespaces: development and production.

Let's create two new namespaces to hold our work.

Use the file namespace-dev.yaml which describes a development namespace:

admin/namespace-dev.yaml 
apiVersion: v1
kind: Namespace
metadata:
  name: development
  labels:
    name: development

Create the development namespace using kubectl.

kubectl create -f https://k8s.io/examples/admin/namespace-dev.yaml

Save the following contents into file namespace-prod.yaml which describes a production namespace:

admin/namespace-prod.yaml 
apiVersion: v1
kind: Namespace
metadata:
  name: production
  labels:
    name: production

And then let's create the production namespace using kubectl.

kubectl create -f https://k8s.io/examples/admin/namespace-prod.yaml

To be sure things are right, let's list all of the namespaces in our cluster.

kubectl get namespaces --show-labels

NAME          STATUS    AGE       LABELS
default       Active    32m       <none>
development   Active    29s       name=development
production    Active    23s       name=production

Create pods in each namespace

A Kubernetes namespace provides the scope for Pods, Services, and Deployments in the cluster.

Users interacting with one namespace do not see the content in another namespace.

To demonstrate this, let's spin up a simple Deployment and Pods in the development namespace.

We first check what is the current context:

kubectl config view

apiVersion: v1
clusters:
- cluster:
    certificate-authority-data: REDACTED
    server: https://130.211.122.180
  name: lithe-cocoa-92103_kubernetes
contexts:
- context:
    cluster: lithe-cocoa-92103_kubernetes
    user: lithe-cocoa-92103_kubernetes
  name: lithe-cocoa-92103_kubernetes
current-context: lithe-cocoa-92103_kubernetes
kind: Config
preferences: {}
users:
- name: lithe-cocoa-92103_kubernetes
  user:
    client-certificate-data: REDACTED
    client-key-data: REDACTED
    token: 65rZW78y8HbwXXtSXuUw9DbP4FLjHi4b
- name: lithe-cocoa-92103_kubernetes-basic-auth
  user:
    password: h5M0FtUUIflBSdI7
    username: admin

kubectl config current-context

lithe-cocoa-92103_kubernetes

The next step is to define a context for the kubectl client to work in each namespace. The value of "cluster" and "user" fields are copied from the current context.

kubectl config set-context dev --namespace=development \
  --cluster=lithe-cocoa-92103_kubernetes \
  --user=lithe-cocoa-92103_kubernetes

kubectl config set-context prod --namespace=production \
  --cluster=lithe-cocoa-92103_kubernetes \
  --user=lithe-cocoa-92103_kubernetes

By default, the above commands add two contexts that are saved into file .kube/config. You can now view the contexts and alternate against the two new request contexts depending on which namespace you wish to work against.

To view the new contexts:

kubectl config view

apiVersion: v1
clusters:
- cluster:
    certificate-authority-data: REDACTED
    server: https://130.211.122.180
  name: lithe-cocoa-92103_kubernetes
contexts:
- context:
    cluster: lithe-cocoa-92103_kubernetes
    user: lithe-cocoa-92103_kubernetes
  name: lithe-cocoa-92103_kubernetes
- context:
    cluster: lithe-cocoa-92103_kubernetes
    namespace: development
    user: lithe-cocoa-92103_kubernetes
  name: dev
- context:
    cluster: lithe-cocoa-92103_kubernetes
    namespace: production
    user: lithe-cocoa-92103_kubernetes
  name: prod
current-context: lithe-cocoa-92103_kubernetes
kind: Config
preferences: {}
users:
- name: lithe-cocoa-92103_kubernetes
  user:
    client-certificate-data: REDACTED
    client-key-data: REDACTED
    token: 65rZW78y8HbwXXtSXuUw9DbP4FLjHi4b
- name: lithe-cocoa-92103_kubernetes-basic-auth
  user:
    password: h5M0FtUUIflBSdI7
    username: admin

Let's switch to operate in the development namespace.

kubectl config use-context dev

You can verify your current context by doing the following:

kubectl config current-context

dev

At this point, all requests we make to the Kubernetes cluster from the command line are scoped to the development namespace.

Let's create some contents.

admin/snowflake-deployment.yaml 
apiVersion: apps/v1
kind: Deployment
metadata:
  labels:
    app: snowflake
  name: snowflake
spec:
  replicas: 2
  selector:
    matchLabels:
      app: snowflake
  template:
    metadata:
      labels:
        app: snowflake
    spec:
      containers:
      - image: registry.k8s.io/serve_hostname
        imagePullPolicy: Always
        name: snowflake

Apply the manifest to create a Deployment

kubectl apply -f https://k8s.io/examples/admin/snowflake-deployment.yaml

We have created a deployment whose replica size is 2 that is running the pod called snowflake with a basic container that serves the hostname.

kubectl get deployment

NAME         READY   UP-TO-DATE   AVAILABLE   AGE
snowflake    2/2     2            2           2m

kubectl get pods -l app=snowflake

NAME                         READY     STATUS    RESTARTS   AGE
snowflake-3968820950-9dgr8   1/1       Running   0          2m
snowflake-3968820950-vgc4n   1/1       Running   0          2m

And this is great, developers are able to do what they want, and they do not have to worry about affecting content in the production namespace.

Let's switch to the production namespace and show how resources in one namespace are hidden from the other.

kubectl config use-context prod

The production namespace should be empty, and the following commands should return nothing.

kubectl get deployment
kubectl get pods

Production likes to run cattle, so let's create some cattle pods.

kubectl create deployment cattle --image=registry.k8s.io/serve_hostname --replicas=5

kubectl get deployment

NAME         READY   UP-TO-DATE   AVAILABLE   AGE
cattle       5/5     5            5           10s

kubectl get pods -l app=cattle

NAME                      READY     STATUS    RESTARTS   AGE
cattle-2263376956-41xy6   1/1       Running   0          34s
cattle-2263376956-kw466   1/1       Running   0          34s
cattle-2263376956-n4v97   1/1       Running   0          34s
cattle-2263376956-p5p3i   1/1       Running   0          34s
cattle-2263376956-sxpth   1/1       Running   0          34s

At this point, it should be clear that the resources users create in one namespace are hidden from the other namespace.

As the policy support in Kubernetes evolves, we will extend this scenario to show how you can provide different authorization rules for each namespace.

31 - Operating etcd clusters for Kubernetes

etcd is a consistent and highly-available key value store used as Kubernetes' backing store for all cluster data.

If your Kubernetes cluster uses etcd as its backing store, make sure you have a back up plan for the data.

You can find in-depth information about etcd in the official documentation.

Before you begin

Before you follow steps in this page to deploy, manage, back up or restore etcd, you need to understand the typical expectations for operating an etcd cluster. Refer to the etcd documentation for more context.

Key details include:

  • The minimum recommended etcd versions to run in production are 3.4.22+ and 3.5.6+.

  • etcd is a leader-based distributed system. Ensure that the leader periodically send heartbeats on time to all followers to keep the cluster stable.

  • You should run etcd as a cluster with an odd number of members.

  • Aim to ensure that no resource starvation occurs.

    Performance and stability of the cluster is sensitive to network and disk I/O. Any resource starvation can lead to heartbeat timeout, causing instability of the cluster. An unstable etcd indicates that no leader is elected. Under such circumstances, a cluster cannot make any changes to its current state, which implies no new pods can be scheduled.

Resource requirements for etcd

Operating etcd with limited resources is suitable only for testing purposes. For deploying in production, advanced hardware configuration is required. Before deploying etcd in production, see resource requirement reference.

Keeping etcd clusters stable is critical to the stability of Kubernetes clusters. Therefore, run etcd clusters on dedicated machines or isolated environments for guaranteed resource requirements.

Tools

Depending on which specific outcome you're working on, you will need the etcdctl tool or the etcdutl tool (you may need both).

Understanding etcdctl and etcdutl

etcdctl and etcdutl are command-line tools used to interact with etcd clusters, but they serve different purposes:

  • etcdctl: This is the primary command-line client for interacting with etcd over a network. It is used for day-to-day operations such as managing keys and values, administering the cluster, checking health, and more.

  • etcdutl: This is an administration utility designed to operate directly on etcd data files, including migrating data between etcd versions, defragmenting the database, restoring snapshots, and validating data consistency. For network operations, etcdctl should be used.

For more information on etcdutl, you can refer to the etcd recovery documentation.

Starting etcd clusters

This section covers starting a single-node and multi-node etcd cluster.

This guide assumes that etcd is already installed.

Single-node etcd cluster

Use a single-node etcd cluster only for testing purposes.

  1. Run the following:

    etcd --listen-client-urls=http://$PRIVATE_IP:2379 \
   --advertise-client-urls=http://$PRIVATE_IP:2379

  2. Start the Kubernetes API server with the flag --etcd-servers=$PRIVATE_IP:2379.

    Make sure PRIVATE_IP is set to your etcd client IP.

Multi-node etcd cluster

For durability and high availability, run etcd as a multi-node cluster in production and back it up periodically. A five-member cluster is recommended in production. For more information, see FAQ documentation.

As you're using Kubernetes, you have the option to run etcd as a container inside one or more Pods. The kubeadm tool sets up etcd static pods by default, or you can deploy a separate cluster and instruct kubeadm to use that etcd cluster as the control plane's backing store.

You configure an etcd cluster either by static member information or by dynamic discovery. For more information on clustering, see etcd clustering documentation.

For an example, consider a five-member etcd cluster running with the following client URLs: http://$IP1:2379, http://$IP2:2379, http://$IP3:2379, http://$IP4:2379, and http://$IP5:2379. To start a Kubernetes API server:

  1. Run the following:

    etcd --listen-client-urls=http://$IP1:2379,http://$IP2:2379,http://$IP3:2379,http://$IP4:2379,http://$IP5:2379 --advertise-client-urls=http://$IP1:2379,http://$IP2:2379,http://$IP3:2379,http://$IP4:2379,http://$IP5:2379

  2. Start the Kubernetes API servers with the flag --etcd-servers=$IP1:2379,$IP2:2379,$IP3:2379,$IP4:2379,$IP5:2379.

    Make sure the IP<n> variables are set to your client IP addresses.

Multi-node etcd cluster with load balancer

To run a load balancing etcd cluster:

  1. Set up an etcd cluster.
  2. Configure a load balancer in front of the etcd cluster. For example, let the address of the load balancer be $LB.
  3. Start Kubernetes API Servers with the flag --etcd-servers=$LB:2379.

Securing etcd clusters

Access to etcd is equivalent to root permission in the cluster so ideally only the API server should have access to it. Considering the sensitivity of the data, it is recommended to grant permission to only those nodes that require access to etcd clusters.

To secure etcd, either set up firewall rules or use the security features provided by etcd. etcd security features depend on x509 Public Key Infrastructure (PKI). To begin, establish secure communication channels by generating a key and certificate pair. For example, use key pairs peer.key and peer.cert for securing communication between etcd members, and client.key and client.cert for securing communication between etcd and its clients. See the example scripts provided by the etcd project to generate key pairs and CA files for client authentication.

Securing communication

To configure etcd with secure peer communication, specify flags --peer-key-file=peer.key and --peer-cert-file=peer.cert, and use HTTPS as the URL schema.

Similarly, to configure etcd with secure client communication, specify flags --key-file=k8sclient.key and --cert-file=k8sclient.cert, and use HTTPS as the URL schema. Here is an example on a client command that uses secure communication:

ETCDCTL_API=3 etcdctl --endpoints 10.2.0.9:2379 \
  --cert=/etc/kubernetes/pki/etcd/server.crt \
  --key=/etc/kubernetes/pki/etcd/server.key \
  --cacert=/etc/kubernetes/pki/etcd/ca.crt \
  member list

Limiting access of etcd clusters

After configuring secure communication, restrict the access of the etcd cluster to only the Kubernetes API servers using TLS authentication.

For example, consider key pairs k8sclient.key and k8sclient.cert that are trusted by the CA etcd.ca. When etcd is configured with --client-cert-auth along with TLS, it verifies the certificates from clients by using system CAs or the CA passed in by --trusted-ca-file flag. Specifying flags --client-cert-auth=true and --trusted-ca-file=etcd.ca will restrict the access to clients with the certificate k8sclient.cert.

Once etcd is configured correctly, only clients with valid certificates can access it. To give Kubernetes API servers the access, configure them with the flags --etcd-certfile=k8sclient.cert, --etcd-keyfile=k8sclient.key and --etcd-cafile=ca.cert.

Replacing a failed etcd member

etcd cluster achieves high availability by tolerating minor member failures. However, to improve the overall health of the cluster, replace failed members immediately. When multiple members fail, replace them one by one. Replacing a failed member involves two steps: removing the failed member and adding a new member.

Though etcd keeps unique member IDs internally, it is recommended to use a unique name for each member to avoid human errors. For example, consider a three-member etcd cluster. Let the URLs be, member1=http://10.0.0.1, member2=http://10.0.0.2, and member3=http://10.0.0.3. When member1 fails, replace it with member4=http://10.0.0.4.

  1. Get the member ID of the failed member1:

    etcdctl --endpoints=http://10.0.0.2,http://10.0.0.3 member list

    The following message is displayed:

    8211f1d0f64f3269, started, member1, http://10.0.0.1:2380, http://10.0.0.1:2379
91bc3c398fb3c146, started, member2, http://10.0.0.2:2380, http://10.0.0.2:2379
fd422379fda50e48, started, member3, http://10.0.0.3:2380, http://10.0.0.3:2379

  2. Do either of the following:

    1. If each Kubernetes API server is configured to communicate with all etcd members, remove the failed member from the --etcd-servers flag, then restart each Kubernetes API server.
    2. If each Kubernetes API server communicates with a single etcd member, then stop the Kubernetes API server that communicates with the failed etcd.

  3. Stop the etcd server on the broken node. It is possible that other clients besides the Kubernetes API server are causing traffic to etcd and it is desirable to stop all traffic to prevent writes to the data directory.

  4. Remove the failed member:

    etcdctl member remove 8211f1d0f64f3269

    The following message is displayed:

    Removed member 8211f1d0f64f3269 from cluster

  5. Add the new member:

    etcdctl member add member4 --peer-urls=http://10.0.0.4:2380

    The following message is displayed:

    Member 2be1eb8f84b7f63e added to cluster ef37ad9dc622a7c4

  6. Start the newly added member on a machine with the IP 10.0.0.4:

    export ETCD_NAME="member4"
export ETCD_INITIAL_CLUSTER="member2=http://10.0.0.2:2380,member3=http://10.0.0.3:2380,member4=http://10.0.0.4:2380"
export ETCD_INITIAL_CLUSTER_STATE=existing
etcd [flags]

  7. Do either of the following:

    1. If each Kubernetes API server is configured to communicate with all etcd members, add the newly added member to the --etcd-servers flag, then restart each Kubernetes API server.
    2. If each Kubernetes API server communicates with a single etcd member, start the Kubernetes API server that was stopped in step 2. Then configure Kubernetes API server clients to again route requests to the Kubernetes API server that was stopped. This can often be done by configuring a load balancer.

For more information on cluster reconfiguration, see etcd reconfiguration documentation.

Backing up an etcd cluster

All Kubernetes objects are stored in etcd. Periodically backing up the etcd cluster data is important to recover Kubernetes clusters under disaster scenarios, such as losing all control plane nodes. The snapshot file contains all the Kubernetes state and critical information. In order to keep the sensitive Kubernetes data safe, encrypt the snapshot files.

Backing up an etcd cluster can be accomplished in two ways: etcd built-in snapshot and volume snapshot.

Built-in snapshot

etcd supports built-in snapshot. A snapshot may either be created from a live member with the etcdctl snapshot save command or by copying the member/snap/db file from an etcd data directory that is not currently used by an etcd process. Creating the snapshot will not affect the performance of the member.

Below is an example for creating a snapshot of the keyspace served by $ENDPOINT to the file snapshot.db:

ETCDCTL_API=3 etcdctl --endpoints $ENDPOINT snapshot save snapshot.db

Verify the snapshot:

The below example depicts the usage of the etcdutl tool for verifying a snapshot:

etcdutl --write-out=table snapshot status snapshot.db

This should generate an output resembling the example provided below:

+----------+----------+------------+------------+
|   HASH   | REVISION | TOTAL KEYS | TOTAL SIZE |
+----------+----------+------------+------------+
| fe01cf57 |       10 |          7 | 2.1 MB     |
+----------+----------+------------+------------+

The below example depicts the usage of the etcdctl tool for verifying a snapshot:

export ETCDCTL_API=3
etcdctl --write-out=table snapshot status snapshot.db

This should generate an output resembling the example provided below:

Deprecated: Use `etcdutl snapshot status` instead.

+----------+----------+------------+------------+
|   HASH   | REVISION | TOTAL KEYS | TOTAL SIZE |
+----------+----------+------------+------------+
| fe01cf57 |       10 |          7 | 2.1 MB     |
+----------+----------+------------+------------+

Volume snapshot

If etcd is running on a storage volume that supports backup, such as Amazon Elastic Block Store, back up etcd data by creating a snapshot of the storage volume.

Snapshot using etcdctl options

We can also create the snapshot using various options given by etcdctl. For example:

ETCDCTL_API=3 etcdctl -h

will list various options available from etcdctl. For example, you can create a snapshot by specifying the endpoint, certificates and key as shown below:

ETCDCTL_API=3 etcdctl --endpoints=https://127.0.0.1:2379 \
  --cacert=<trusted-ca-file> --cert=<cert-file> --key=<key-file> \
  snapshot save <backup-file-location>

where trusted-ca-file, cert-file and key-file can be obtained from the description of the etcd Pod.

Scaling out etcd clusters

Scaling out etcd clusters increases availability by trading off performance. Scaling does not increase cluster performance nor capability. A general rule is not to scale out or in etcd clusters. Do not configure any auto scaling groups for etcd clusters. It is strongly recommended to always run a static five-member etcd cluster for production Kubernetes clusters at any officially supported scale.

A reasonable scaling is to upgrade a three-member cluster to a five-member one, when more reliability is desired. See etcd reconfiguration documentation for information on how to add members into an existing cluster.

Restoring an etcd cluster

etcd supports restoring from snapshots that are taken from an etcd process of the major.minor version. Restoring a version from a different patch version of etcd is also supported. A restore operation is employed to recover the data of a failed cluster.

Before starting the restore operation, a snapshot file must be present. It can either be a snapshot file from a previous backup operation, or from a remaining data directory.

When restoring the cluster using etcdutl, use the --data-dir option to specify to which folder the cluster should be restored:

etcdutl --data-dir <data-dir-location> snapshot restore snapshot.db

where <data-dir-location> is a directory that will be created during the restore process.

The below example depicts the usage of the etcdctl tool for the restore operation:

export ETCDCTL_API=3
etcdctl --data-dir <data-dir-location> snapshot restore snapshot.db

If <data-dir-location> is the same folder as before, delete it and stop the etcd process before restoring the cluster. Otherwise, change etcd configuration and restart the etcd process after restoration to have it use the new data directory: first change /etc/kubernetes/manifests/etcd.yaml's volumes.hostPath.path for name: etcd-data to <data-dir-location>, then execute kubectl -n kube-system delete pod <name-of-etcd-pod> or systemctl restart kubelet.service (or both).

For more information and examples on restoring a cluster from a snapshot file, see etcd disaster recovery documentation.

If the access URLs of the restored cluster are changed from the previous cluster, the Kubernetes API server must be reconfigured accordingly. In this case, restart Kubernetes API servers with the flag --etcd-servers=$NEW_ETCD_CLUSTER instead of the flag --etcd-servers=$OLD_ETCD_CLUSTER. Replace $NEW_ETCD_CLUSTER and $OLD_ETCD_CLUSTER with the respective IP addresses. If a load balancer is used in front of an etcd cluster, you might need to update the load balancer instead.

If the majority of etcd members have permanently failed, the etcd cluster is considered failed. In this scenario, Kubernetes cannot make any changes to its current state. Although the scheduled pods might continue to run, no new pods can be scheduled. In such cases, recover the etcd cluster and potentially reconfigure Kubernetes API servers to fix the issue.

Upgrading etcd clusters

For details on etcd upgrade, refer to the etcd upgrades documentation.

Maintaining etcd clusters

For more details on etcd maintenance, please refer to the etcd maintenance documentation.

Cluster defragmentation

Defragmentation is an expensive operation, so it should be executed as infrequently as possible. On the other hand, it's also necessary to make sure any etcd member will not exceed the storage quota. The Kubernetes project recommends that when you perform defragmentation, you use a tool such as etcd-defrag.

You can also run the defragmentation tool as a Kubernetes CronJob, to make sure that defragmentation happens regularly. See etcd-defrag-cronjob.yaml for details.

32 - Reserve Compute Resources for System Daemons

Kubernetes nodes can be scheduled to Capacity. Pods can consume all the available capacity on a node by default. This is an issue because nodes typically run quite a few system daemons that power the OS and Kubernetes itself. Unless resources are set aside for these system daemons, pods and system daemons compete for resources and lead to resource starvation issues on the node.

The kubelet exposes a feature named 'Node Allocatable' that helps to reserve compute resources for system daemons. Kubernetes recommends cluster administrators to configure 'Node Allocatable' based on their workload density on each node.

Before you begin

You need to have a Kubernetes cluster, and the kubectl command-line tool must be configured to communicate with your cluster. It is recommended to run this tutorial on a cluster with at least two nodes that are not acting as control plane hosts. If you do not already have a cluster, you can create one by using minikube or you can use one of these Kubernetes playgrounds:

You can configure below kubelet configuration settings using the kubelet configuration file.

Node Allocatable

node capacity

'Allocatable' on a Kubernetes node is defined as the amount of compute resources that are available for pods. The scheduler does not over-subscribe 'Allocatable'. 'CPU', 'memory' and 'ephemeral-storage' are supported as of now.

Node Allocatable is exposed as part of v1.Node object in the API and as part of kubectl describe node in the CLI.

Resources can be reserved for two categories of system daemons in the kubelet.

Enabling QoS and Pod level cgroups

To properly enforce node allocatable constraints on the node, you must enable the new cgroup hierarchy via the cgroupsPerQOS setting. This setting is enabled by default. When enabled, the kubelet will parent all end-user pods under a cgroup hierarchy managed by the kubelet.

Configuring a cgroup driver

The kubelet supports manipulation of the cgroup hierarchy on the host using a cgroup driver. The driver is configured via the cgroupDriver setting.

The supported values are the following:

  • cgroupfs is the default driver that performs direct manipulation of the cgroup filesystem on the host in order to manage cgroup sandboxes.
  • systemd is an alternative driver that manages cgroup sandboxes using transient slices for resources that are supported by that init system.

Depending on the configuration of the associated container runtime, operators may have to choose a particular cgroup driver to ensure proper system behavior. For example, if operators use the systemd cgroup driver provided by the containerd runtime, the kubelet must be configured to use the systemd cgroup driver.

Kube Reserved

  • KubeletConfiguration Setting: kubeReserved: {}. Example value {cpu: 100m, memory: 100Mi, ephemeral-storage: 1Gi, pid=1000}
  • KubeletConfiguration Setting: kubeReservedCgroup: ""

kubeReserved is meant to capture resource reservation for kubernetes system daemons like the kubelet, container runtime, etc. It is not meant to reserve resources for system daemons that are run as pods. kubeReserved is typically a function of pod density on the nodes.

In addition to cpu, memory, and ephemeral-storage, pid may be specified to reserve the specified number of process IDs for kubernetes system daemons.

To optionally enforce kubeReserved on kubernetes system daemons, specify the parent control group for kube daemons as the value for kubeReservedCgroup setting, and add kube-reserved to enforceNodeAllocatable.

It is recommended that the kubernetes system daemons are placed under a top level control group (runtime.slice on systemd machines for example). Each system daemon should ideally run within its own child control group. Refer to the design proposal for more details on recommended control group hierarchy.

Note that Kubelet does not create kubeReservedCgroup if it doesn't exist. The kubelet will fail to start if an invalid cgroup is specified. With systemd cgroup driver, you should follow a specific pattern for the name of the cgroup you define: the name should be the value you set for kubeReservedCgroup, with .slice appended.

System Reserved

  • KubeletConfiguration Setting: systemReserved: {}. Example value {cpu: 100m, memory: 100Mi, ephemeral-storage: 1Gi, pid=1000}
  • KubeletConfiguration Setting: systemReservedCgroup: ""

systemReserved is meant to capture resource reservation for OS system daemons like sshd, udev, etc. systemReserved should reserve memory for the kernel too since kernel memory is not accounted to pods in Kubernetes at this time. Reserving resources for user login sessions is also recommended (user.slice in systemd world).

In addition to cpu, memory, and ephemeral-storage, pid may be specified to reserve the specified number of process IDs for OS system daemons.

To optionally enforce systemReserved on system daemons, specify the parent control group for OS system daemons as the value for systemReservedCgroup setting, and add system-reserved to enforceNodeAllocatable.

It is recommended that the OS system daemons are placed under a top level control group (system.slice on systemd machines for example).

Note that kubelet does not create systemReservedCgroup if it doesn't exist. kubelet will fail if an invalid cgroup is specified. With systemd cgroup driver, you should follow a specific pattern for the name of the cgroup you define: the name should be the value you set for systemReservedCgroup, with .slice appended.

Explicitly Reserved CPU List

FEATURE STATE: Kubernetes v1.17 [stable]

KubeletConfiguration Setting: reservedSystemCPUs:. Example value 0-3

reservedSystemCPUs is meant to define an explicit CPU set for OS system daemons and kubernetes system daemons. reservedSystemCPUs is for systems that do not intend to define separate top level cgroups for OS system daemons and kubernetes system daemons with regard to cpuset resource. If the Kubelet does not have kubeReservedCgroup and systemReservedCgroup, the explicit cpuset provided by reservedSystemCPUs will take precedence over the CPUs defined by kubeReservedCgroup and systemReservedCgroup options.

This option is specifically designed for Telco/NFV use cases where uncontrolled interrupts/timers may impact the workload performance. you can use this option to define the explicit cpuset for the system/kubernetes daemons as well as the interrupts/timers, so the rest CPUs on the system can be used exclusively for workloads, with less impact from uncontrolled interrupts/timers. To move the system daemon, kubernetes daemons and interrupts/timers to the explicit cpuset defined by this option, other mechanism outside Kubernetes should be used. For example: in Centos, you can do this using the tuned toolset.

Eviction Thresholds

KubeletConfiguration Setting: evictionHard: {memory.available: "100Mi", nodefs.available: "10%", nodefs.inodesFree: "5%", imagefs.available: "15%"}. Example value: {memory.available: "<500Mi"}

Memory pressure at the node level leads to System OOMs which affects the entire node and all pods running on it. Nodes can go offline temporarily until memory has been reclaimed. To avoid (or reduce the probability of) system OOMs kubelet provides out of resource management. Evictions are supported for memory and ephemeral-storage only. By reserving some memory via evictionHard setting, the kubelet attempts to evict pods whenever memory availability on the node drops below the reserved value. Hypothetically, if system daemons did not exist on a node, pods cannot use more than capacity - eviction-hard. For this reason, resources reserved for evictions are not available for pods.

Enforcing Node Allocatable

KubeletConfiguration setting: enforceNodeAllocatable: [pods]. Example value: [pods,system-reserved,kube-reserved]

The scheduler treats 'Allocatable' as the available capacity for pods.

kubelet enforce 'Allocatable' across pods by default. Enforcement is performed by evicting pods whenever the overall usage across all pods exceeds 'Allocatable'. More details on eviction policy can be found on the node pressure eviction page. This enforcement is controlled by specifying pods value to the KubeletConfiguration setting enforceNodeAllocatable.

Optionally, kubelet can be made to enforce kubeReserved and systemReserved by specifying kube-reserved & system-reserved values in the same setting. Note that to enforce kubeReserved or systemReserved, kubeReservedCgroup or systemReservedCgroup needs to be specified respectively.

General Guidelines

System daemons are expected to be treated similar to Guaranteed pods. System daemons can burst within their bounding control groups and this behavior needs to be managed as part of kubernetes deployments. For example, kubelet should have its own control group and share kubeReserved resources with the container runtime. However, Kubelet cannot burst and use up all available Node resources if kubeReserved is enforced.

Be extra careful while enforcing systemReserved reservation since it can lead to critical system services being CPU starved, OOM killed, or unable to fork on the node. The recommendation is to enforce systemReserved only if a user has profiled their nodes exhaustively to come up with precise estimates and is confident in their ability to recover if any process in that group is oom-killed.

  • To begin with enforce 'Allocatable' on pods.
  • Once adequate monitoring and alerting is in place to track kube system daemons, attempt to enforce kubeReserved based on usage heuristics.
  • If absolutely necessary, enforce systemReserved over time.

The resource requirements of kube system daemons may grow over time as more and more features are added. Over time, kubernetes project will attempt to bring down utilization of node system daemons, but that is not a priority as of now. So expect a drop in Allocatable capacity in future releases.

Example Scenario

Here is an example to illustrate Node Allocatable computation:

  • Node has 32Gi of memory, 16 CPUs and 100Gi of Storage
  • kubeReserved is set to {cpu: 1000m, memory: 2Gi, ephemeral-storage: 1Gi}
  • systemReserved is set to {cpu: 500m, memory: 1Gi, ephemeral-storage: 1Gi}
  • evictionHard is set to {memory.available: "<500Mi", nodefs.available: "<10%"}

Under this scenario, 'Allocatable' will be 14.5 CPUs, 28.5Gi of memory and 88Gi of local storage. Scheduler ensures that the total memory requests across all pods on this node does not exceed 28.5Gi and storage doesn't exceed 88Gi. Kubelet evicts pods whenever the overall memory usage across pods exceeds 28.5Gi, or if overall disk usage exceeds 88Gi. If all processes on the node consume as much CPU as they can, pods together cannot consume more than 14.5 CPUs.

If kubeReserved and/or systemReserved is not enforced and system daemons exceed their reservation, kubelet evicts pods whenever the overall node memory usage is higher than 31.5Gi or storage is greater than 90Gi.

33 - Running Kubernetes Node Components as a Non-root User

FEATURE STATE: Kubernetes v1.22 [alpha]

This document describes how to run Kubernetes Node components such as kubelet, CRI, OCI, and CNI without root privileges, by using a user namespace.

This technique is also known as rootless mode.

Before you begin

Your Kubernetes server must be at or later than version 1.22. To check the version, enter kubectl version.

Running Kubernetes inside Rootless Docker/Podman

kind

kind supports running Kubernetes inside Rootless Docker or Rootless Podman.

See Running kind with Rootless Docker.

minikube

minikube also supports running Kubernetes inside Rootless Docker or Rootless Podman.

See the Minikube documentation:

Running Kubernetes inside Unprivileged Containers

sysbox

Sysbox is an open-source container runtime (similar to "runc") that supports running system-level workloads such as Docker and Kubernetes inside unprivileged containers isolated with the Linux user namespace.

See Sysbox Quick Start Guide: Kubernetes-in-Docker for more info.

Sysbox supports running Kubernetes inside unprivileged containers without requiring Cgroup v2 and without the KubeletInUserNamespace feature gate. It does this by exposing specially crafted /proc and /sys filesystems inside the container plus several other advanced OS virtualization techniques.

Running Rootless Kubernetes directly on a host

K3s

K3s experimentally supports rootless mode.

See Running K3s with Rootless mode for the usage.

Usernetes

Usernetes is a reference distribution of Kubernetes that can be installed under $HOME directory without the root privilege.

Usernetes supports both containerd and CRI-O as CRI runtimes. Usernetes supports multi-node clusters using Flannel (VXLAN).

See the Usernetes repo for the usage.

Manually deploy a node that runs the kubelet in a user namespace

This section provides hints for running Kubernetes in a user namespace manually.

Creating a user namespace

The first step is to create a user namespace.

If you are trying to run Kubernetes in a user-namespaced container such as Rootless Docker/Podman or LXC/LXD, you are all set, and you can go to the next subsection.

Otherwise you have to create a user namespace by yourself, by calling unshare(2) with CLONE_NEWUSER.

A user namespace can be also unshared by using command line tools such as:

After unsharing the user namespace, you will also have to unshare other namespaces such as mount namespace.

You do not need to call chroot() nor pivot_root() after unsharing the mount namespace, however, you have to mount writable filesystems on several directories in the namespace.

At least, the following directories need to be writable in the namespace (not outside the namespace):

  • /etc
  • /run
  • /var/logs
  • /var/lib/kubelet
  • /var/lib/cni
  • /var/lib/containerd (for containerd)
  • /var/lib/containers (for CRI-O)

Creating a delegated cgroup tree

In addition to the user namespace, you also need to have a writable cgroup tree with cgroup v2.

If you are trying to run Kubernetes in Rootless Docker/Podman or LXC/LXD on a systemd-based host, you are all set.

Otherwise you have to create a systemd unit with Delegate=yes property to delegate a cgroup tree with writable permission.

On your node, systemd must already be configured to allow delegation; for more details, see cgroup v2 in the Rootless Containers documentation.

Configuring network

The network namespace of the Node components has to have a non-loopback interface, which can be for example configured with slirp4netns, VPNKit, or lxc-user-nic(1).

The network namespaces of the Pods can be configured with regular CNI plugins. For multi-node networking, Flannel (VXLAN, 8472/UDP) is known to work.

Ports such as the kubelet port (10250/TCP) and NodePort service ports have to be exposed from the Node network namespace to the host with an external port forwarder, such as RootlessKit, slirp4netns, or socat(1).

You can use the port forwarder from K3s. See Running K3s in Rootless Mode for more details. The implementation can be found in the pkg/rootlessports package of k3s.

Configuring CRI

The kubelet relies on a container runtime. You should deploy a container runtime such as containerd or CRI-O and ensure that it is running within the user namespace before the kubelet starts.

Running CRI plugin of containerd in a user namespace is supported since containerd 1.4.

Running containerd within a user namespace requires the following configurations.

version = 2

[plugins."io.containerd.grpc.v1.cri"]
# Disable AppArmor
  disable_apparmor = true
# Ignore an error during setting oom_score_adj
  restrict_oom_score_adj = true
# Disable hugetlb cgroup v2 controller (because systemd does not support delegating hugetlb controller)
  disable_hugetlb_controller = true

[plugins."io.containerd.grpc.v1.cri".containerd]
# Using non-fuse overlayfs is also possible for kernel >= 5.11, but requires SELinux to be disabled
  snapshotter = "fuse-overlayfs"

[plugins."io.containerd.grpc.v1.cri".containerd.runtimes.runc.options]
# We use cgroupfs that is delegated by systemd, so we do not use SystemdCgroup driver
# (unless you run another systemd in the namespace)
  SystemdCgroup = false

The default path of the configuration file is /etc/containerd/config.toml. The path can be specified with containerd -c /path/to/containerd/config.toml.

Running CRI-O in a user namespace is supported since CRI-O 1.22.

CRI-O requires an environment variable _CRIO_ROOTLESS=1 to be set.

The following configurations are also recommended:

[crio]
  storage_driver = "overlay"
# Using non-fuse overlayfs is also possible for kernel >= 5.11, but requires SELinux to be disabled
  storage_option = ["overlay.mount_program=/usr/local/bin/fuse-overlayfs"]

[crio.runtime]
# We use cgroupfs that is delegated by systemd, so we do not use "systemd" driver
# (unless you run another systemd in the namespace)
  cgroup_manager = "cgroupfs"

The default path of the configuration file is /etc/crio/crio.conf. The path can be specified with crio --config /path/to/crio/crio.conf.

Configuring kubelet

Running kubelet in a user namespace requires the following configuration:

apiVersion: kubelet.config.k8s.io/v1beta1
kind: KubeletConfiguration
featureGates:
  KubeletInUserNamespace: true
# We use cgroupfs that is delegated by systemd, so we do not use "systemd" driver
# (unless you run another systemd in the namespace)
cgroupDriver: "cgroupfs"

When the KubeletInUserNamespace feature gate is enabled, the kubelet ignores errors that may happen during setting the following sysctl values on the node.

  • vm.overcommit_memory
  • vm.panic_on_oom
  • kernel.panic
  • kernel.panic_on_oops
  • kernel.keys.root_maxkeys
  • kernel.keys.root_maxbytes.

Within a user namespace, the kubelet also ignores any error raised from trying to open /dev/kmsg. This feature gate also allows kube-proxy to ignore an error during setting RLIMIT_NOFILE.

The KubeletInUserNamespace feature gate was introduced in Kubernetes v1.22 with "alpha" status.

Running kubelet in a user namespace without using this feature gate is also possible by mounting a specially crafted proc filesystem (as done by Sysbox), but not officially supported.

Configuring kube-proxy

Running kube-proxy in a user namespace requires the following configuration:

apiVersion: kubeproxy.config.k8s.io/v1alpha1
kind: KubeProxyConfiguration
mode: "iptables" # or "userspace"
conntrack:
# Skip setting sysctl value "net.netfilter.nf_conntrack_max"
  maxPerCore: 0
# Skip setting "net.netfilter.nf_conntrack_tcp_timeout_established"
  tcpEstablishedTimeout: 0s
# Skip setting "net.netfilter.nf_conntrack_tcp_timeout_close"
  tcpCloseWaitTimeout: 0s

Caveats

  • Most of "non-local" volume drivers such as nfs and iscsi do not work. Local volumes like local, hostPath, emptyDir, configMap, secret, and downwardAPI are known to work.

  • Some CNI plugins may not work. Flannel (VXLAN) is known to work.

For more on this, see the Caveats and Future work page on the rootlesscontaine.rs website.

See Also

34 - Safely Drain a Node

This page shows how to safely drain a node, optionally respecting the PodDisruptionBudget you have defined.

Before you begin

This task assumes that you have met the following prerequisites:

  1. You do not require your applications to be highly available during the node drain, or
  2. You have read about the PodDisruptionBudget concept, and have configured PodDisruptionBudgets for applications that need them.

(Optional) Configure a disruption budget

To ensure that your workloads remain available during maintenance, you can configure a PodDisruptionBudget.

If availability is important for any applications that run or could run on the node(s) that you are draining, configure a PodDisruptionBudgets first and then continue following this guide.

It is recommended to set AlwaysAllow Unhealthy Pod Eviction Policy to your PodDisruptionBudgets to support eviction of misbehaving applications during a node drain. The default behavior is to wait for the application pods to become healthy before the drain can proceed.

Use kubectl drain to remove a node from service

You can use kubectl drain to safely evict all of your pods from a node before you perform maintenance on the node (e.g. kernel upgrade, hardware maintenance, etc.). Safe evictions allow the pod's containers to gracefully terminate and will respect the PodDisruptionBudgets you have specified.

When kubectl drain returns successfully, that indicates that all of the pods (except the ones excluded as described in the previous paragraph) have been safely evicted (respecting the desired graceful termination period, and respecting the PodDisruptionBudget you have defined). It is then safe to bring down the node by powering down its physical machine or, if running on a cloud platform, deleting its virtual machine.

First, identify the name of the node you wish to drain. You can list all of the nodes in your cluster with

kubectl get nodes

Next, tell Kubernetes to drain the node:

kubectl drain --ignore-daemonsets <node name>

If there are pods managed by a DaemonSet, you will need to specify --ignore-daemonsets with kubectl to successfully drain the node. The kubectl drain subcommand on its own does not actually drain a node of its DaemonSet pods: the DaemonSet controller (part of the control plane) immediately replaces missing Pods with new equivalent Pods. The DaemonSet controller also creates Pods that ignore unschedulable taints, which allows the new Pods to launch onto a node that you are draining.

Once it returns (without giving an error), you can power down the node (or equivalently, if on a cloud platform, delete the virtual machine backing the node). If you leave the node in the cluster during the maintenance operation, you need to run

kubectl uncordon <node name>

afterwards to tell Kubernetes that it can resume scheduling new pods onto the node.

Draining multiple nodes in parallel

The kubectl drain command should only be issued to a single node at a time. However, you can run multiple kubectl drain commands for different nodes in parallel, in different terminals or in the background. Multiple drain commands running concurrently will still respect the PodDisruptionBudget you specify.

For example, if you have a StatefulSet with three replicas and have set a PodDisruptionBudget for that set specifying minAvailable: 2, kubectl drain only evicts a pod from the StatefulSet if all three replicas pods are healthy; if then you issue multiple drain commands in parallel, Kubernetes respects the PodDisruptionBudget and ensures that only 1 (calculated as replicas - minAvailable) Pod is unavailable at any given time. Any drains that would cause the number of healthy replicas to fall below the specified budget are blocked.

The Eviction API

If you prefer not to use kubectl drain (such as to avoid calling to an external command, or to get finer control over the pod eviction process), you can also programmatically cause evictions using the eviction API.

For more information, see API-initiated eviction.

What's next

35 - Securing a Cluster

This document covers topics related to protecting a cluster from accidental or malicious access and provides recommendations on overall security.

Before you begin

  • You need to have a Kubernetes cluster, and the kubectl command-line tool must be configured to communicate with your cluster. It is recommended to run this tutorial on a cluster with at least two nodes that are not acting as control plane hosts. If you do not already have a cluster, you can create one by using minikube or you can use one of these Kubernetes playgrounds:

    To check the version, enter kubectl version.

Controlling access to the Kubernetes API

As Kubernetes is entirely API-driven, controlling and limiting who can access the cluster and what actions they are allowed to perform is the first line of defense.

Use Transport Layer Security (TLS) for all API traffic

Kubernetes expects that all API communication in the cluster is encrypted by default with TLS, and the majority of installation methods will allow the necessary certificates to be created and distributed to the cluster components. Note that some components and installation methods may enable local ports over HTTP and administrators should familiarize themselves with the settings of each component to identify potentially unsecured traffic.

API Authentication

Choose an authentication mechanism for the API servers to use that matches the common access patterns when you install a cluster. For instance, small, single-user clusters may wish to use a simple certificate or static Bearer token approach. Larger clusters may wish to integrate an existing OIDC or LDAP server that allow users to be subdivided into groups.

All API clients must be authenticated, even those that are part of the infrastructure like nodes, proxies, the scheduler, and volume plugins. These clients are typically service accounts or use x509 client certificates, and they are created automatically at cluster startup or are setup as part of the cluster installation.

Consult the authentication reference document for more information.

API Authorization

Once authenticated, every API call is also expected to pass an authorization check. Kubernetes ships an integrated Role-Based Access Control (RBAC) component that matches an incoming user or group to a set of permissions bundled into roles. These permissions combine verbs (get, create, delete) with resources (pods, services, nodes) and can be namespace-scoped or cluster-scoped. A set of out-of-the-box roles are provided that offer reasonable default separation of responsibility depending on what actions a client might want to perform. It is recommended that you use the Node and RBAC authorizers together, in combination with the NodeRestriction admission plugin.

As with authentication, simple and broad roles may be appropriate for smaller clusters, but as more users interact with the cluster, it may become necessary to separate teams into separate namespaces with more limited roles.

With authorization, it is important to understand how updates on one object may cause actions in other places. For instance, a user may not be able to create pods directly, but allowing them to create a deployment, which creates pods on their behalf, will let them create those pods indirectly. Likewise, deleting a node from the API will result in the pods scheduled to that node being terminated and recreated on other nodes. The out-of-the box roles represent a balance between flexibility and common use cases, but more limited roles should be carefully reviewed to prevent accidental escalation. You can make roles specific to your use case if the out-of-box ones don't meet your needs.

Consult the authorization reference section for more information.

Controlling access to the Kubelet

Kubelets expose HTTPS endpoints which grant powerful control over the node and containers. By default Kubelets allow unauthenticated access to this API.

Production clusters should enable Kubelet authentication and authorization.

Consult the Kubelet authentication/authorization reference for more information.

Controlling the capabilities of a workload or user at runtime

Authorization in Kubernetes is intentionally high level, focused on coarse actions on resources. More powerful controls exist as policies to limit by use case how those objects act on the cluster, themselves, and other resources.

Limiting resource usage on a cluster

Resource quota limits the number or capacity of resources granted to a namespace. This is most often used to limit the amount of CPU, memory, or persistent disk a namespace can allocate, but can also control how many pods, services, or volumes exist in each namespace.

Limit ranges restrict the maximum or minimum size of some of the resources above, to prevent users from requesting unreasonably high or low values for commonly reserved resources like memory, or to provide default limits when none are specified.

Controlling what privileges containers run with

A pod definition contains a security context that allows it to request access to run as a specific Linux user on a node (like root), access to run privileged or access the host network, and other controls that would otherwise allow it to run unfettered on a hosting node.

You can configure Pod security admission to enforce use of a particular Pod Security Standard in a namespace, or to detect breaches.

Generally, most application workloads need limited access to host resources so they can successfully run as a root process (uid 0) without access to host information. However, considering the privileges associated with the root user, you should write application containers to run as a non-root user. Similarly, administrators who wish to prevent client applications from escaping their containers should apply the Baseline or Restricted Pod Security Standard.

Preventing containers from loading unwanted kernel modules

The Linux kernel automatically loads kernel modules from disk if needed in certain circumstances, such as when a piece of hardware is attached or a filesystem is mounted. Of particular relevance to Kubernetes, even unprivileged processes can cause certain network-protocol-related kernel modules to be loaded, just by creating a socket of the appropriate type. This may allow an attacker to exploit a security hole in a kernel module that the administrator assumed was not in use.

To prevent specific modules from being automatically loaded, you can uninstall them from the node, or add rules to block them. On most Linux distributions, you can do that by creating a file such as /etc/modprobe.d/kubernetes-blacklist.conf with contents like:

# DCCP is unlikely to be needed, has had multiple serious
# vulnerabilities, and is not well-maintained.
blacklist dccp

# SCTP is not used in most Kubernetes clusters, and has also had
# vulnerabilities in the past.
blacklist sctp

To block module loading more generically, you can use a Linux Security Module (such as SELinux) to completely deny the module_request permission to containers, preventing the kernel from loading modules for containers under any circumstances. (Pods would still be able to use modules that had been loaded manually, or modules that were loaded by the kernel on behalf of some more-privileged process.)

Restricting network access

The network policies for a namespace allows application authors to restrict which pods in other namespaces may access pods and ports within their namespaces. Many of the supported Kubernetes networking providers now respect network policy.

Quota and limit ranges can also be used to control whether users may request node ports or load-balanced services, which on many clusters can control whether those users applications are visible outside of the cluster.

Additional protections may be available that control network rules on a per-plugin or per- environment basis, such as per-node firewalls, physically separating cluster nodes to prevent cross talk, or advanced networking policy.

Restricting cloud metadata API access

Cloud platforms (AWS, Azure, GCE, etc.) often expose metadata services locally to instances. By default these APIs are accessible by pods running on an instance and can contain cloud credentials for that node, or provisioning data such as kubelet credentials. These credentials can be used to escalate within the cluster or to other cloud services under the same account.

When running Kubernetes on a cloud platform, limit permissions given to instance credentials, use network policies to restrict pod access to the metadata API, and avoid using provisioning data to deliver secrets.

Controlling which nodes pods may access

By default, there are no restrictions on which nodes may run a pod. Kubernetes offers a rich set of policies for controlling placement of pods onto nodes and the taint-based pod placement and eviction that are available to end users. For many clusters use of these policies to separate workloads can be a convention that authors adopt or enforce via tooling.

As an administrator, a beta admission plugin PodNodeSelector can be used to force pods within a namespace to default or require a specific node selector, and if end users cannot alter namespaces, this can strongly limit the placement of all of the pods in a specific workload.

Protecting cluster components from compromise

This section describes some common patterns for protecting clusters from compromise.

Restrict access to etcd

Write access to the etcd backend for the API is equivalent to gaining root on the entire cluster, and read access can be used to escalate fairly quickly. Administrators should always use strong credentials from the API servers to their etcd server, such as mutual auth via TLS client certificates, and it is often recommended to isolate the etcd servers behind a firewall that only the API servers may access.

Enable audit logging

The audit logger is a beta feature that records actions taken by the API for later analysis in the event of a compromise. It is recommended to enable audit logging and archive the audit file on a secure server.

Restrict access to alpha or beta features

Alpha and beta Kubernetes features are in active development and may have limitations or bugs that result in security vulnerabilities. Always assess the value an alpha or beta feature may provide against the possible risk to your security posture. When in doubt, disable features you do not use.

Rotate infrastructure credentials frequently

The shorter the lifetime of a secret or credential the harder it is for an attacker to make use of that credential. Set short lifetimes on certificates and automate their rotation. Use an authentication provider that can control how long issued tokens are available and use short lifetimes where possible. If you use service-account tokens in external integrations, plan to rotate those tokens frequently. For example, once the bootstrap phase is complete, a bootstrap token used for setting up nodes should be revoked or its authorization removed.

Review third party integrations before enabling them

Many third party integrations to Kubernetes may alter the security profile of your cluster. When enabling an integration, always review the permissions that an extension requests before granting it access. For example, many security integrations may request access to view all secrets on your cluster which is effectively making that component a cluster admin. When in doubt, restrict the integration to functioning in a single namespace if possible.

Components that create pods may also be unexpectedly powerful if they can do so inside namespaces like the kube-system namespace, because those pods can gain access to service account secrets or run with elevated permissions if those service accounts are granted access to permissive PodSecurityPolicies.

If you use Pod Security admission and allow any component to create Pods within a namespace that permits privileged Pods, those Pods may be able to escape their containers and use this widened access to elevate their privileges.

You should not allow untrusted components to create Pods in any system namespace (those with names that start with kube-) nor in any namespace where that access grant allows the possibility of privilege escalation.

Encrypt secrets at rest

In general, the etcd database will contain any information accessible via the Kubernetes API and may grant an attacker significant visibility into the state of your cluster. Always encrypt your backups using a well reviewed backup and encryption solution, and consider using full disk encryption where possible.

Kubernetes supports optional encryption at rest for information in the Kubernetes API. This lets you ensure that when Kubernetes stores data for objects (for example, Secret or ConfigMap objects), the API server writes an encrypted representation of the object. That encryption means that even someone who has access to etcd backup data is unable to view the content of those objects. In Kubernetes 1.32 you can also encrypt custom resources; encryption-at-rest for extension APIs defined in CustomResourceDefinitions was added to Kubernetes as part of the v1.26 release.

Receiving alerts for security updates and reporting vulnerabilities

Join the kubernetes-announce group for emails about security announcements. See the security reporting page for more on how to report vulnerabilities.

What's next

36 - Set Kubelet Parameters Via A Configuration File

Before you begin

Some steps in this page use the jq tool. If you don't have jq, you can install it via your operating system's software sources, or fetch it from https://jqlang.github.io/jq/.

Some steps also involve installing curl, which can be installed via your operating system's software sources.

A subset of the kubelet's configuration parameters may be set via an on-disk config file, as a substitute for command-line flags.

Providing parameters via a config file is the recommended approach because it simplifies node deployment and configuration management.

Create the config file

The subset of the kubelet's configuration that can be configured via a file is defined by the KubeletConfiguration struct.

The configuration file must be a JSON or YAML representation of the parameters in this struct. Make sure the kubelet has read permissions on the file.

Here is an example of what this file might look like:

apiVersion: kubelet.config.k8s.io/v1beta1
kind: KubeletConfiguration
address: "192.168.0.8"
port: 20250
serializeImagePulls: false
evictionHard:
    memory.available:  "100Mi"
    nodefs.available:  "10%"
    nodefs.inodesFree: "5%"
    imagefs.available: "15%"

In this example, the kubelet is configured with the following settings:

  1. address: The kubelet will serve on IP address 192.168.0.8.

  2. port: The kubelet will serve on port 20250.

  3. serializeImagePulls: Image pulls will be done in parallel.

  4. evictionHard: The kubelet will evict Pods under one of the following conditions:

    • When the node's available memory drops below 100MiB.
    • When the node's main filesystem's available space is less than 10%.
    • When the image filesystem's available space is less than 15%.
    • When more than 95% of the node's main filesystem's inodes are in use.

The imagefs is an optional filesystem that container runtimes use to store container images and container writable layers.

Start a kubelet process configured via the config file

Start the kubelet with the --config flag set to the path of the kubelet's config file. The kubelet will then load its config from this file.

Note that command line flags which target the same value as a config file will override that value. This helps ensure backwards compatibility with the command-line API.

Note that relative file paths in the kubelet config file are resolved relative to the location of the kubelet config file, whereas relative paths in command line flags are resolved relative to the kubelet's current working directory.

Note that some default values differ between command-line flags and the kubelet config file. If --config is provided and the values are not specified via the command line, the defaults for the KubeletConfiguration version apply. In the above example, this version is kubelet.config.k8s.io/v1beta1.

Drop-in directory for kubelet configuration files

FEATURE STATE: Kubernetes v1.30 [beta]

You can specify a drop-in configuration directory for the kubelet. By default, the kubelet does not look for drop-in configuration files anywhere - you must specify a path. For example: --config-dir=/etc/kubernetes/kubelet.conf.d

For Kubernetes v1.28 to v1.29, you can only specify --config-dir if you also set the environment variable KUBELET_CONFIG_DROPIN_DIR_ALPHA for the kubelet process (the value of that variable does not matter).

The kubelet processes files in its config drop-in directory by sorting the entire file name alphanumerically. For instance, 00-kubelet.conf is processed first, and then overridden with a file named 01-kubelet.conf.

These files may contain partial configurations but should not be invalid and must include type metadata, specifically apiVersion and kind. Validation is only performed on the final resulting configuration structure stored internally in the kubelet. This offers flexibility in managing and merging kubelet configurations from different sources while preventing undesirable configurations. However, it is important to note that behavior varies based on the data type of the configuration fields.

Different data types in the kubelet configuration structure merge differently. See the reference document for more information.

Kubelet configuration merging order

On startup, the kubelet merges configuration from:

  • Feature gates specified over the command line (lowest precedence).
  • The kubelet configuration.
  • Drop-in configuration files, according to sort order.
  • Command line arguments excluding feature gates (highest precedence).

Viewing the kubelet configuration

Since the configuration could now be spread over multiple files with this feature, if someone wants to inspect the final actuated configuration, they can follow these steps to inspect the kubelet configuration:

  1. Start a proxy server using kubectl proxy in your terminal.

    kubectl proxy

    Which gives output like:

    Starting to serve on 127.0.0.1:8001

  2. Open another terminal window and use curl to fetch the kubelet configuration. Replace <node-name> with the actual name of your node:

    curl -X GET http://127.0.0.1:8001/api/v1/nodes/<node-name>/proxy/configz | jq .

    {
  "kubeletconfig": {
    "enableServer": true,
    "staticPodPath": "/var/run/kubernetes/static-pods",
    "syncFrequency": "1m0s",
    "fileCheckFrequency": "20s",
    "httpCheckFrequency": "20s",
    "address": "192.168.1.16",
    "port": 10250,
    "readOnlyPort": 10255,
    "tlsCertFile": "/var/lib/kubelet/pki/kubelet.crt",
    "tlsPrivateKeyFile": "/var/lib/kubelet/pki/kubelet.key",
    "rotateCertificates": true,
    "authentication": {
      "x509": {
        "clientCAFile": "/var/run/kubernetes/client-ca.crt"
      },
      "webhook": {
        "enabled": true,
        "cacheTTL": "2m0s"
      },
      "anonymous": {
        "enabled": true
      }
    },
    "authorization": {
      "mode": "AlwaysAllow",
      "webhook": {
        "cacheAuthorizedTTL": "5m0s",
        "cacheUnauthorizedTTL": "30s"
      }
    },
    "registryPullQPS": 5,
    "registryBurst": 10,
    "eventRecordQPS": 50,
    "eventBurst": 100,
    "enableDebuggingHandlers": true,
    "healthzPort": 10248,
    "healthzBindAddress": "127.0.0.1",
    "oomScoreAdj": -999,
    "clusterDomain": "cluster.local",
    "clusterDNS": [
      "10.0.0.10"
    ],
    "streamingConnectionIdleTimeout": "4h0m0s",
    "nodeStatusUpdateFrequency": "10s",
    "nodeStatusReportFrequency": "5m0s",
    "nodeLeaseDurationSeconds": 40,
    "imageMinimumGCAge": "2m0s",
    "imageMaximumGCAge": "0s",
    "imageGCHighThresholdPercent": 85,
    "imageGCLowThresholdPercent": 80,
    "volumeStatsAggPeriod": "1m0s",
    "cgroupsPerQOS": true,
    "cgroupDriver": "systemd",
    "cpuManagerPolicy": "none",
    "cpuManagerReconcilePeriod": "10s",
    "memoryManagerPolicy": "None",
    "topologyManagerPolicy": "none",
    "topologyManagerScope": "container",
    "runtimeRequestTimeout": "2m0s",
    "hairpinMode": "promiscuous-bridge",
    "maxPods": 110,
    "podPidsLimit": -1,
    "resolvConf": "/run/systemd/resolve/resolv.conf",
    "cpuCFSQuota": true,
    "cpuCFSQuotaPeriod": "100ms",
    "nodeStatusMaxImages": 50,
    "maxOpenFiles": 1000000,
    "contentType": "application/vnd.kubernetes.protobuf",
    "kubeAPIQPS": 50,
    "kubeAPIBurst": 100,
    "serializeImagePulls": true,
    "evictionHard": {
      "imagefs.available": "15%",
      "memory.available": "100Mi",
      "nodefs.available": "10%",
      "nodefs.inodesFree": "5%"
    },
    "evictionPressureTransitionPeriod": "1m0s",
    "enableControllerAttachDetach": true,
    "makeIPTablesUtilChains": true,
    "iptablesMasqueradeBit": 14,
    "iptablesDropBit": 15,
    "featureGates": {
      "AllAlpha": false
    },
    "failSwapOn": false,
    "memorySwap": {},
    "containerLogMaxSize": "10Mi",
    "containerLogMaxFiles": 5,
    "configMapAndSecretChangeDetectionStrategy": "Watch",
    "enforceNodeAllocatable": [
      "pods"
    ],
    "volumePluginDir": "/usr/libexec/kubernetes/kubelet-plugins/volume/exec/",
    "logging": {
      "format": "text",
      "flushFrequency": "5s",
      "verbosity": 3,
      "options": {
        "json": {
          "infoBufferSize": "0"
        }
      }
    },
    "enableSystemLogHandler": true,
    "enableSystemLogQuery": false,
    "shutdownGracePeriod": "0s",
    "shutdownGracePeriodCriticalPods": "0s",
    "enableProfilingHandler": true,
    "enableDebugFlagsHandler": true,
    "seccompDefault": false,
    "memoryThrottlingFactor": 0.9,
    "registerNode": true,
    "localStorageCapacityIsolation": true,
    "containerRuntimeEndpoint": "unix:///var/run/crio/crio.sock"
  }
}

What's next

37 - Share a Cluster with Namespaces

This page shows how to view, work in, and delete namespaces. The page also shows how to use Kubernetes namespaces to subdivide your cluster.

Before you begin

Viewing namespaces

List the current namespaces in a cluster using:

kubectl get namespaces

NAME              STATUS   AGE
default           Active   11d
kube-node-lease   Active   11d
kube-public       Active   11d
kube-system       Active   11d

Kubernetes starts with four initial namespaces:

  • default The default namespace for objects with no other namespace
  • kube-node-lease This namespace holds Lease objects associated with each node. Node leases allow the kubelet to send heartbeats so that the control plane can detect node failure.
  • kube-public This namespace is created automatically and is readable by all users (including those not authenticated). This namespace is mostly reserved for cluster usage, in case that some resources should be visible and readable publicly throughout the whole cluster. The public aspect of this namespace is only a convention, not a requirement.
  • kube-system The namespace for objects created by the Kubernetes system

You can also get the summary of a specific namespace using:

kubectl get namespaces <name>

Or you can get detailed information with:

kubectl describe namespaces <name>

Name:           default
Labels:         <none>
Annotations:    <none>
Status:         Active

No resource quota.

Resource Limits
 Type       Resource    Min Max Default
 ----               --------    --- --- ---
 Container          cpu         -   -   100m

Note that these details show both resource quota (if present) as well as resource limit ranges.

Resource quota tracks aggregate usage of resources in the Namespace and allows cluster operators to define Hard resource usage limits that a Namespace may consume.

A limit range defines min/max constraints on the amount of resources a single entity can consume in a Namespace.

See Admission control: Limit Range

A namespace can be in one of two phases:

  • Active the namespace is in use
  • Terminating the namespace is being deleted, and can not be used for new objects

For more details, see Namespace in the API reference.

Creating a new namespace

Create a new YAML file called my-namespace.yaml with the contents:

apiVersion: v1
kind: Namespace
metadata:
  name: <insert-namespace-name-here>

Then run:

kubectl create -f ./my-namespace.yaml

Alternatively, you can create namespace using below command:

kubectl create namespace <insert-namespace-name-here>

The name of your namespace must be a valid DNS label.

There's an optional field finalizers, which allows observables to purge resources whenever the namespace is deleted. Keep in mind that if you specify a nonexistent finalizer, the namespace will be created but will get stuck in the Terminating state if the user tries to delete it.

More information on finalizers can be found in the namespace design doc.

Deleting a namespace

Delete a namespace with

kubectl delete namespaces <insert-some-namespace-name>

This delete is asynchronous, so for a time you will see the namespace in the Terminating state.

Subdividing your cluster using Kubernetes namespaces

By default, a Kubernetes cluster will instantiate a default namespace when provisioning the cluster to hold the default set of Pods, Services, and Deployments used by the cluster.

Assuming you have a fresh cluster, you can introspect the available namespaces by doing the following:

kubectl get namespaces

NAME      STATUS    AGE
default   Active    13m

Create new namespaces

For this exercise, we will create two additional Kubernetes namespaces to hold our content.

In a scenario where an organization is using a shared Kubernetes cluster for development and production use cases:

  • The development team would like to maintain a space in the cluster where they can get a view on the list of Pods, Services, and Deployments they use to build and run their application. In this space, Kubernetes resources come and go, and the restrictions on who can or cannot modify resources are relaxed to enable agile development.

  • The operations team would like to maintain a space in the cluster where they can enforce strict procedures on who can or cannot manipulate the set of Pods, Services, and Deployments that run the production site.

One pattern this organization could follow is to partition the Kubernetes cluster into two namespaces: development and production. Let's create two new namespaces to hold our work.

Create the development namespace using kubectl:

kubectl create -f https://k8s.io/examples/admin/namespace-dev.json

And then let's create the production namespace using kubectl:

kubectl create -f https://k8s.io/examples/admin/namespace-prod.json

To be sure things are right, list all of the namespaces in our cluster.

kubectl get namespaces --show-labels

NAME          STATUS    AGE       LABELS
default       Active    32m       <none>
development   Active    29s       name=development
production    Active    23s       name=production

Create pods in each namespace

A Kubernetes namespace provides the scope for Pods, Services, and Deployments in the cluster. Users interacting with one namespace do not see the content in another namespace. To demonstrate this, let's spin up a simple Deployment and Pods in the development namespace.

kubectl create deployment snowflake \
  --image=registry.k8s.io/serve_hostname \
  -n=development --replicas=2

We have created a deployment whose replica size is 2 that is running the pod called snowflake with a basic container that serves the hostname.

kubectl get deployment -n=development

NAME         READY   UP-TO-DATE   AVAILABLE   AGE
snowflake    2/2     2            2           2m

kubectl get pods -l app=snowflake -n=development

NAME                         READY     STATUS    RESTARTS   AGE
snowflake-3968820950-9dgr8   1/1       Running   0          2m
snowflake-3968820950-vgc4n   1/1       Running   0          2m

And this is great, developers are able to do what they want, and they do not have to worry about affecting content in the production namespace.

Let's switch to the production namespace and show how resources in one namespace are hidden from the other. The production namespace should be empty, and the following commands should return nothing.

kubectl get deployment -n=production
kubectl get pods -n=production

Production likes to run cattle, so let's create some cattle pods.

kubectl create deployment cattle --image=registry.k8s.io/serve_hostname -n=production
kubectl scale deployment cattle --replicas=5 -n=production

kubectl get deployment -n=production

NAME         READY   UP-TO-DATE   AVAILABLE   AGE
cattle       5/5     5            5           10s

kubectl get pods -l app=cattle -n=production

NAME                      READY     STATUS    RESTARTS   AGE
cattle-2263376956-41xy6   1/1       Running   0          34s
cattle-2263376956-kw466   1/1       Running   0          34s
cattle-2263376956-n4v97   1/1       Running   0          34s
cattle-2263376956-p5p3i   1/1       Running   0          34s
cattle-2263376956-sxpth   1/1       Running   0          34s

At this point, it should be clear that the resources users create in one namespace are hidden from the other namespace.

As the policy support in Kubernetes evolves, we will extend this scenario to show how you can provide different authorization rules for each namespace.

Understanding the motivation for using namespaces

A single cluster should be able to satisfy the needs of multiple users or groups of users (henceforth in this document a user community).

Kubernetes namespaces help different projects, teams, or customers to share a Kubernetes cluster.

It does this by providing the following:

  1. A scope for names.
  2. A mechanism to attach authorization and policy to a subsection of the cluster.

Use of multiple namespaces is optional.

Each user community wants to be able to work in isolation from other communities. Each user community has its own:

  1. resources (pods, services, replication controllers, etc.)
  2. policies (who can or cannot perform actions in their community)
  3. constraints (this community is allowed this much quota, etc.)

A cluster operator may create a Namespace for each unique user community.

The Namespace provides a unique scope for:

  1. named resources (to avoid basic naming collisions)
  2. delegated management authority to trusted users
  3. ability to limit community resource consumption

Use cases include:

  1. As a cluster operator, I want to support multiple user communities on a single cluster.
  2. As a cluster operator, I want to delegate authority to partitions of the cluster to trusted users in those communities.
  3. As a cluster operator, I want to limit the amount of resources each community can consume in order to limit the impact to other communities using the cluster.
  4. As a cluster user, I want to interact with resources that are pertinent to my user community in isolation of what other user communities are doing on the cluster.

Understanding namespaces and DNS

When you create a Service, it creates a corresponding DNS entry. This entry is of the form <service-name>.<namespace-name>.svc.cluster.local, which means that if a container uses <service-name> it will resolve to the service which is local to a namespace. This is useful for using the same configuration across multiple namespaces such as Development, Staging and Production. If you want to reach across namespaces, you need to use the fully qualified domain name (FQDN).

What's next

38 - Upgrade A Cluster

This page provides an overview of the steps you should follow to upgrade a Kubernetes cluster.

The Kubernetes project recommends upgrading to the latest patch releases promptly, and to ensure that you are running a supported minor release of Kubernetes. Following this recommendation helps you to to stay secure.

The way that you upgrade a cluster depends on how you initially deployed it and on any subsequent changes.

At a high level, the steps you perform are:

  • Upgrade the control plane
  • Upgrade the nodes in your cluster
  • Upgrade clients such as kubectl
  • Adjust manifests and other resources based on the API changes that accompany the new Kubernetes version

Before you begin

You must have an existing cluster. This page is about upgrading from Kubernetes 1.31 to Kubernetes 1.32. If your cluster is not currently running Kubernetes 1.31 then please check the documentation for the version of Kubernetes that you plan to upgrade to.

Upgrade approaches

kubeadm

If your cluster was deployed using the kubeadm tool, refer to Upgrading kubeadm clusters for detailed information on how to upgrade the cluster.

Once you have upgraded the cluster, remember to install the latest version of kubectl.

Manual deployments

You should manually update the control plane following this sequence:

  • etcd (all instances)
  • kube-apiserver (all control plane hosts)
  • kube-controller-manager
  • kube-scheduler
  • cloud controller manager, if you use one

At this point you should install the latest version of kubectl.

For each node in your cluster, drain that node and then either replace it with a new node that uses the 1.32 kubelet, or upgrade the kubelet on that node and bring the node back into service.

Other deployments

Refer to the documentation for your cluster deployment tool to learn the recommended set up steps for maintenance.

Post-upgrade tasks

Switch your cluster's storage API version

The objects that are serialized into etcd for a cluster's internal representation of the Kubernetes resources active in the cluster are written using a particular version of the API.

When the supported API changes, these objects may need to be rewritten in the newer API. Failure to do this will eventually result in resources that are no longer decodable or usable by the Kubernetes API server.

For each affected object, fetch it using the latest supported API and then write it back also using the latest supported API.

Update manifests

Upgrading to a new Kubernetes version can provide new APIs.

You can use kubectl convert command to convert manifests between different API versions. For example:

kubectl convert -f pod.yaml --output-version v1

The kubectl tool replaces the contents of pod.yaml with a manifest that sets kind to Pod (unchanged), but with a revised apiVersion.

Device Plugins

If your cluster is running device plugins and the node needs to be upgraded to a Kubernetes release with a newer device plugin API version, device plugins must be upgraded to support both version before the node is upgraded in order to guarantee that device allocations continue to complete successfully during the upgrade.

Refer to API compatibility and Kubelet Device Manager API Versions for more details.

39 - Use Cascading Deletion in a Cluster

This page shows you how to specify the type of cascading deletion to use in your cluster during garbage collection.

Before you begin

You need to have a Kubernetes cluster, and the kubectl command-line tool must be configured to communicate with your cluster. It is recommended to run this tutorial on a cluster with at least two nodes that are not acting as control plane hosts. If you do not already have a cluster, you can create one by using minikube or you can use one of these Kubernetes playgrounds:

You also need to create a sample Deployment to experiment with the different types of cascading deletion. You will need to recreate the Deployment for each type.

Check owner references on your pods

Check that the ownerReferences field is present on your pods:

kubectl get pods -l app=nginx --output=yaml

The output has an ownerReferences field similar to this:

apiVersion: v1
    ...
    ownerReferences:
    - apiVersion: apps/v1
      blockOwnerDeletion: true
      controller: true
      kind: ReplicaSet
      name: nginx-deployment-6b474476c4
      uid: 4fdcd81c-bd5d-41f7-97af-3a3b759af9a7
    ...

Use foreground cascading deletion

By default, Kubernetes uses background cascading deletion to delete dependents of an object. You can switch to foreground cascading deletion using either kubectl or the Kubernetes API, depending on the Kubernetes version your cluster runs. To check the version, enter kubectl version.

You can delete objects using foreground cascading deletion using kubectl or the Kubernetes API.

Using kubectl

Run the following command:

kubectl delete deployment nginx-deployment --cascade=foreground

Using the Kubernetes API

  1. Start a local proxy session:

    kubectl proxy --port=8080

  2. Use curl to trigger deletion:

    curl -X DELETE localhost:8080/apis/apps/v1/namespaces/default/deployments/nginx-deployment \
    -d '{"kind":"DeleteOptions","apiVersion":"v1","propagationPolicy":"Foreground"}' \
    -H "Content-Type: application/json"

    The output contains a foregroundDeletion finalizer like this:

    "kind": "Deployment",
"apiVersion": "apps/v1",
"metadata": {
    "name": "nginx-deployment",
    "namespace": "default",
    "uid": "d1ce1b02-cae8-4288-8a53-30e84d8fa505",
    "resourceVersion": "1363097",
    "creationTimestamp": "2021-07-08T20:24:37Z",
    "deletionTimestamp": "2021-07-08T20:27:39Z",
    "finalizers": [
      "foregroundDeletion"
    ]
    ...

Use background cascading deletion

  1. Create a sample Deployment.
  2. Use either kubectl or the Kubernetes API to delete the Deployment, depending on the Kubernetes version your cluster runs. To check the version, enter kubectl version.

You can delete objects using background cascading deletion using kubectl or the Kubernetes API.

Kubernetes uses background cascading deletion by default, and does so even if you run the following commands without the --cascade flag or the propagationPolicy argument.

Using kubectl

Run the following command:

kubectl delete deployment nginx-deployment --cascade=background

Using the Kubernetes API

  1. Start a local proxy session:

    kubectl proxy --port=8080

  2. Use curl to trigger deletion:

    curl -X DELETE localhost:8080/apis/apps/v1/namespaces/default/deployments/nginx-deployment \
    -d '{"kind":"DeleteOptions","apiVersion":"v1","propagationPolicy":"Background"}' \
    -H "Content-Type: application/json"

    The output is similar to this:

    "kind": "Status",
"apiVersion": "v1",
...
"status": "Success",
"details": {
    "name": "nginx-deployment",
    "group": "apps",
    "kind": "deployments",
    "uid": "cc9eefb9-2d49-4445-b1c1-d261c9396456"
}

Delete owner objects and orphan dependents

By default, when you tell Kubernetes to delete an object, the controller also deletes dependent objects. You can make Kubernetes orphan these dependents using kubectl or the Kubernetes API, depending on the Kubernetes version your cluster runs. To check the version, enter kubectl version.

Using kubectl

Run the following command:

kubectl delete deployment nginx-deployment --cascade=orphan

Using the Kubernetes API

  1. Start a local proxy session:

    kubectl proxy --port=8080

  2. Use curl to trigger deletion:

    curl -X DELETE localhost:8080/apis/apps/v1/namespaces/default/deployments/nginx-deployment \
    -d '{"kind":"DeleteOptions","apiVersion":"v1","propagationPolicy":"Orphan"}' \
    -H "Content-Type: application/json"

    The output contains orphan in the finalizers field, similar to this:

    "kind": "Deployment",
"apiVersion": "apps/v1",
"namespace": "default",
"uid": "6f577034-42a0-479d-be21-78018c466f1f",
"creationTimestamp": "2021-07-09T16:46:37Z",
"deletionTimestamp": "2021-07-09T16:47:08Z",
"deletionGracePeriodSeconds": 0,
"finalizers": [
  "orphan"
],
...

You can check that the Pods managed by the Deployment are still running:

kubectl get pods -l app=nginx

What's next

40 - Using a KMS provider for data encryption

This page shows how to configure a Key Management Service (KMS) provider and plugin to enable secret data encryption. In Kubernetes 1.32 there are two versions of KMS at-rest encryption. You should use KMS v2 if feasible because KMS v1 is deprecated (since Kubernetes v1.28) and disabled by default (since Kubernetes v1.29). KMS v2 offers significantly better performance characteristics than KMS v1.

Before you begin

You need to have a Kubernetes cluster, and the kubectl command-line tool must be configured to communicate with your cluster. It is recommended to run this tutorial on a cluster with at least two nodes that are not acting as control plane hosts. If you do not already have a cluster, you can create one by using minikube or you can use one of these Kubernetes playgrounds:

The version of Kubernetes that you need depends on which KMS API version you have selected. Kubernetes recommends using KMS v2.

  • If you selected KMS API v1 to support clusters prior to version v1.27 or if you have a legacy KMS plugin that only supports KMS v1, any supported Kubernetes version will work. This API is deprecated as of Kubernetes v1.28. Kubernetes does not recommend the use of this API.
To check the version, enter kubectl version.

KMS v1

FEATURE STATE: Kubernetes v1.28 [deprecated]

  • Kubernetes version 1.10.0 or later is required

  • For version 1.29 and later, the v1 implementation of KMS is disabled by default. To enable the feature, set --feature-gates=KMSv1=true to configure a KMS v1 provider.

  • Your cluster must use etcd v3 or later

KMS v2

FEATURE STATE: Kubernetes v1.29 [stable]
  • Your cluster must use etcd v3 or later

KMS encryption and per-object encryption keys

The KMS encryption provider uses an envelope encryption scheme to encrypt data in etcd. The data is encrypted using a data encryption key (DEK). The DEKs are encrypted with a key encryption key (KEK) that is stored and managed in a remote KMS.

If you use the (deprecated) v1 implementation of KMS, a new DEK is generated for each encryption.

With KMS v2, a new DEK is generated per encryption: the API server uses a key derivation function to generate single use data encryption keys from a secret seed combined with some random data. The seed is rotated whenever the KEK is rotated (see the Understanding key_id and Key Rotation section below for more details).

The KMS provider uses gRPC to communicate with a specific KMS plugin over a UNIX domain socket. The KMS plugin, which is implemented as a gRPC server and deployed on the same host(s) as the Kubernetes control plane, is responsible for all communication with the remote KMS.

Configuring the KMS provider

To configure a KMS provider on the API server, include a provider of type kms in the providers array in the encryption configuration file and set the following properties:

KMS v1

  • apiVersion: API Version for KMS provider. Leave this value empty or set it to v1.
  • name: Display name of the KMS plugin. Cannot be changed once set.
  • endpoint: Listen address of the gRPC server (KMS plugin). The endpoint is a UNIX domain socket.
  • cachesize: Number of data encryption keys (DEKs) to be cached in the clear. When cached, DEKs can be used without another call to the KMS; whereas DEKs that are not cached require a call to the KMS to unwrap.
  • timeout: How long should kube-apiserver wait for kms-plugin to respond before returning an error (default is 3 seconds).

KMS v2

  • apiVersion: API Version for KMS provider. Set this to v2.
  • name: Display name of the KMS plugin. Cannot be changed once set.
  • endpoint: Listen address of the gRPC server (KMS plugin). The endpoint is a UNIX domain socket.
  • timeout: How long should kube-apiserver wait for kms-plugin to respond before returning an error (default is 3 seconds).

KMS v2 does not support the cachesize property. All data encryption keys (DEKs) will be cached in the clear once the server has unwrapped them via a call to the KMS. Once cached, DEKs can be used to perform decryption indefinitely without making a call to the KMS.

See Understanding the encryption at rest configuration.

Implementing a KMS plugin

To implement a KMS plugin, you can develop a new plugin gRPC server or enable a KMS plugin already provided by your cloud provider. You then integrate the plugin with the remote KMS and deploy it on the Kubernetes control plane.

Enabling the KMS supported by your cloud provider

Refer to your cloud provider for instructions on enabling the cloud provider-specific KMS plugin.

Developing a KMS plugin gRPC server

You can develop a KMS plugin gRPC server using a stub file available for Go. For other languages, you use a proto file to create a stub file that you can use to develop the gRPC server code.

KMS v1

  • Using Go: Use the functions and data structures in the stub file: api.pb.go to develop the gRPC server code

  • Using languages other than Go: Use the protoc compiler with the proto file: api.proto to generate a stub file for the specific language

KMS v2

  • Using Go: A high level library is provided to make the process easier. Low level implementations can use the functions and data structures in the stub file: api.pb.go to develop the gRPC server code

  • Using languages other than Go: Use the protoc compiler with the proto file: api.proto to generate a stub file for the specific language

Then use the functions and data structures in the stub file to develop the server code.

Notes

KMS v1

  • kms plugin version: v1beta1

    In response to procedure call Version, a compatible KMS plugin should return v1beta1 as VersionResponse.version.

  • message version: v1beta1

    All messages from KMS provider have the version field set to v1beta1.

  • protocol: UNIX domain socket (unix)

    The plugin is implemented as a gRPC server that listens at UNIX domain socket. The plugin deployment should create a file on the file system to run the gRPC unix domain socket connection. The API server (gRPC client) is configured with the KMS provider (gRPC server) unix domain socket endpoint in order to communicate with it. An abstract Linux socket may be used by starting the endpoint with /@, i.e. unix:///@foo. Care must be taken when using this type of socket as they do not have concept of ACL (unlike traditional file based sockets). However, they are subject to Linux networking namespace, so will only be accessible to containers within the same pod unless host networking is used.

KMS v2

  • KMS plugin version: v2

    In response to the Status remote procedure call, a compatible KMS plugin should return its KMS compatibility version as StatusResponse.version. That status response should also include "ok" as StatusResponse.healthz and a key_id (remote KMS KEK ID) as StatusResponse.key_id. The Kubernetes project recommends you make your plugin compatible with the stable v2 KMS API. Kubernetes 1.32 also supports the v2beta1 API for KMS; future Kubernetes releases are likely to continue supporting that beta version.

    The API server polls the Status procedure call approximately every minute when everything is healthy, and every 10 seconds when the plugin is not healthy. Plugins must take care to optimize this call as it will be under constant load.

  • Encryption

    The EncryptRequest procedure call provides the plaintext and a UID for logging purposes. The response must include the ciphertext, the key_id for the KEK used, and, optionally, any metadata that the KMS plugin needs to aid in future DecryptRequest calls (via the annotations field). The plugin must guarantee that any distinct plaintext results in a distinct response (ciphertext, key_id, annotations).

    If the plugin returns a non-empty annotations map, all map keys must be fully qualified domain names such as example.com. An example use case of annotation is {"kms.example.io/remote-kms-auditid":"<audit ID used by the remote KMS>"}

    The API server does not perform the EncryptRequest procedure call at a high rate. Plugin implementations should still aim to keep each request's latency at under 100 milliseconds.

  • Decryption

    The DecryptRequest procedure call provides the (ciphertext, key_id, annotations) from EncryptRequest and a UID for logging purposes. As expected, it is the inverse of the EncryptRequest call. Plugins must verify that the key_id is one that they understand - they must not attempt to decrypt data unless they are sure that it was encrypted by them at an earlier time.

    The API server may perform thousands of DecryptRequest procedure calls on startup to fill its watch cache. Thus plugin implementations must perform these calls as quickly as possible, and should aim to keep each request's latency at under 10 milliseconds.

  • Understanding key_id and Key Rotation

    The key_id is the public, non-secret name of the remote KMS KEK that is currently in use. It may be logged during regular operation of the API server, and thus must not contain any private data. Plugin implementations are encouraged to use a hash to avoid leaking any data. The KMS v2 metrics take care to hash this value before exposing it via the /metrics endpoint.

    The API server considers the key_id returned from the Status procedure call to be authoritative. Thus, a change to this value signals to the API server that the remote KEK has changed, and data encrypted with the old KEK should be marked stale when a no-op write is performed (as described below). If an EncryptRequest procedure call returns a key_id that is different from Status, the response is thrown away and the plugin is considered unhealthy. Thus implementations must guarantee that the key_id returned from Status will be the same as the one returned by EncryptRequest. Furthermore, plugins must ensure that the key_id is stable and does not flip-flop between values (i.e. during a remote KEK rotation).

    Plugins must not re-use key_ids, even in situations where a previously used remote KEK has been reinstated. For example, if a plugin was using key_id=A, switched to key_id=B, and then went back to key_id=A - instead of reporting key_id=A the plugin should report some derivative value such as key_id=A_001 or use a new value such as key_id=C.

    Since the API server polls Status about every minute, key_id rotation is not immediate. Furthermore, the API server will coast on the last valid state for about three minutes. Thus if a user wants to take a passive approach to storage migration (i.e. by waiting), they must schedule a migration to occur at 3 + N + M minutes after the remote KEK has been rotated (N is how long it takes the plugin to observe the key_id change and M is the desired buffer to allow config changes to be processed - a minimum M of five minutes is recommend). Note that no API server restart is required to perform KEK rotation.

  • protocol: UNIX domain socket (unix)

    The plugin is implemented as a gRPC server that listens at UNIX domain socket. The plugin deployment should create a file on the file system to run the gRPC unix domain socket connection. The API server (gRPC client) is configured with the KMS provider (gRPC server) unix domain socket endpoint in order to communicate with it. An abstract Linux socket may be used by starting the endpoint with /@, i.e. unix:///@foo. Care must be taken when using this type of socket as they do not have concept of ACL (unlike traditional file based sockets). However, they are subject to Linux networking namespace, so will only be accessible to containers within the same pod unless host networking is used.

Integrating a KMS plugin with the remote KMS

The KMS plugin can communicate with the remote KMS using any protocol supported by the KMS. All configuration data, including authentication credentials the KMS plugin uses to communicate with the remote KMS, are stored and managed by the KMS plugin independently. The KMS plugin can encode the ciphertext with additional metadata that may be required before sending it to the KMS for decryption (KMS v2 makes this process easier by providing a dedicated annotations field).

Deploying the KMS plugin

Ensure that the KMS plugin runs on the same host(s) as the Kubernetes API server(s).

Encrypting your data with the KMS provider

To encrypt the data:

  1. Create a new EncryptionConfiguration file using the appropriate properties for the kms provider to encrypt resources like Secrets and ConfigMaps. If you want to encrypt an extension API that is defined in a CustomResourceDefinition, your cluster must be running Kubernetes v1.26 or newer.

  2. Set the --encryption-provider-config flag on the kube-apiserver to point to the location of the configuration file.

  3. --encryption-provider-config-automatic-reload boolean argument determines if the file set by --encryption-provider-config should be automatically reloaded if the disk contents change.

  4. Restart your API server.

KMS v1

apiVersion: apiserver.config.k8s.io/v1
kind: EncryptionConfiguration
resources:
  - resources:
      - secrets
      - configmaps
      - pandas.awesome.bears.example
    providers:
      - kms:
          name: myKmsPluginFoo
          endpoint: unix:///tmp/socketfile-foo.sock
          cachesize: 100
          timeout: 3s
      - kms:
          name: myKmsPluginBar
          endpoint: unix:///tmp/socketfile-bar.sock
          cachesize: 100
          timeout: 3s

KMS v2

apiVersion: apiserver.config.k8s.io/v1
kind: EncryptionConfiguration
resources:
  - resources:
      - secrets
      - configmaps
      - pandas.awesome.bears.example
    providers:
      - kms:
          apiVersion: v2
          name: myKmsPluginFoo
          endpoint: unix:///tmp/socketfile-foo.sock
          timeout: 3s
      - kms:
          apiVersion: v2
          name: myKmsPluginBar
          endpoint: unix:///tmp/socketfile-bar.sock
          timeout: 3s

Setting --encryption-provider-config-automatic-reload to true collapses all health checks to a single health check endpoint. Individual health checks are only available when KMS v1 providers are in use and the encryption config is not auto-reloaded.

The following table summarizes the health check endpoints for each KMS version:

KMS configurations Without Automatic Reload With Automatic Reload
KMS v1 only Individual Healthchecks Single Healthcheck
KMS v2 only Single Healthcheck Single Healthcheck
Both KMS v1 and v2 Individual Healthchecks Single Healthcheck
No KMS None Single Healthcheck

Single Healthcheck means that the only health check endpoint is /healthz/kms-providers.

Individual Healthchecks means that each KMS plugin has an associated health check endpoint based on its location in the encryption config: /healthz/kms-provider-0, /healthz/kms-provider-1 etc.

These healthcheck endpoint paths are hard coded and generated/controlled by the server. The indices for individual healthchecks corresponds to the order in which the KMS encryption config is processed.

Until the steps defined in Ensuring all secrets are encrypted are performed, the providers list should end with the identity: {} provider to allow unencrypted data to be read. Once all resources are encrypted, the identity provider should be removed to prevent the API server from honoring unencrypted data.

For details about the EncryptionConfiguration format, please check the API server encryption API reference.

Verifying that the data is encrypted

When encryption at rest is correctly configured, resources are encrypted on write. After restarting your kube-apiserver, any newly created or updated Secret or other resource types configured in EncryptionConfiguration should be encrypted when stored. To verify, you can use the etcdctl command line program to retrieve the contents of your secret data.

  1. Create a new secret called secret1 in the default namespace:

    kubectl create secret generic secret1 -n default --from-literal=mykey=mydata

  2. Using the etcdctl command line, read that secret out of etcd:

    ETCDCTL_API=3 etcdctl get /kubernetes.io/secrets/default/secret1 [...] | hexdump -C

    where [...] contains the additional arguments for connecting to the etcd server.

  3. Verify the stored secret is prefixed with k8s:enc:kms:v1: for KMS v1 or prefixed with k8s:enc:kms:v2: for KMS v2, which indicates that the kms provider has encrypted the resulting data.

  4. Verify that the secret is correctly decrypted when retrieved via the API:

    kubectl describe secret secret1 -n default

    The Secret should contain mykey: mydata

Ensuring all secrets are encrypted

When encryption at rest is correctly configured, resources are encrypted on write. Thus we can perform an in-place no-op update to ensure that data is encrypted.

The following command reads all secrets and then updates them to apply server side encryption. If an error occurs due to a conflicting write, retry the command. For larger clusters, you may wish to subdivide the secrets by namespace or script an update.

kubectl get secrets --all-namespaces -o json | kubectl replace -f -

Switching from a local encryption provider to the KMS provider

To switch from a local encryption provider to the kms provider and re-encrypt all of the secrets:

  1. Add the kms provider as the first entry in the configuration file as shown in the following example.

    apiVersion: apiserver.config.k8s.io/v1
kind: EncryptionConfiguration
resources:
  - resources:
      - secrets
    providers:
      - kms:
          apiVersion: v2
          name : myKmsPlugin
          endpoint: unix:///tmp/socketfile.sock
      - aescbc:
          keys:
            - name: key1
              secret: <BASE 64 ENCODED SECRET>

  2. Restart all kube-apiserver processes.

  3. Run the following command to force all secrets to be re-encrypted using the kms provider.

    kubectl get secrets --all-namespaces -o json | kubectl replace -f -

What's next

If you no longer want to use encryption for data persisted in the Kubernetes API, read decrypt data that are already stored at rest.

41 - Using CoreDNS for Service Discovery

This page describes the CoreDNS upgrade process and how to install CoreDNS instead of kube-dns.

Before you begin

You need to have a Kubernetes cluster, and the kubectl command-line tool must be configured to communicate with your cluster. It is recommended to run this tutorial on a cluster with at least two nodes that are not acting as control plane hosts. If you do not already have a cluster, you can create one by using minikube or you can use one of these Kubernetes playgrounds:

Your Kubernetes server must be at or later than version v1.9. To check the version, enter kubectl version.

About CoreDNS

CoreDNS is a flexible, extensible DNS server that can serve as the Kubernetes cluster DNS. Like Kubernetes, the CoreDNS project is hosted by the CNCF.

You can use CoreDNS instead of kube-dns in your cluster by replacing kube-dns in an existing deployment, or by using tools like kubeadm that will deploy and upgrade the cluster for you.

Installing CoreDNS

For manual deployment or replacement of kube-dns, see the documentation at the CoreDNS website.

Migrating to CoreDNS

Upgrading an existing cluster with kubeadm

In Kubernetes version 1.21, kubeadm removed its support for kube-dns as a DNS application. For kubeadm v1.32, the only supported cluster DNS application is CoreDNS.

You can move to CoreDNS when you use kubeadm to upgrade a cluster that is using kube-dns. In this case, kubeadm generates the CoreDNS configuration ("Corefile") based upon the kube-dns ConfigMap, preserving configurations for stub domains, and upstream name server.

Upgrading CoreDNS

You can check the version of CoreDNS that kubeadm installs for each version of Kubernetes in the page CoreDNS version in Kubernetes.

CoreDNS can be upgraded manually in case you want to only upgrade CoreDNS or use your own custom image. There is a helpful guideline and walkthrough available to ensure a smooth upgrade. Make sure the existing CoreDNS configuration ("Corefile") is retained when upgrading your cluster.

If you are upgrading your cluster using the kubeadm tool, kubeadm can take care of retaining the existing CoreDNS configuration automatically.

Tuning CoreDNS

When resource utilisation is a concern, it may be useful to tune the configuration of CoreDNS. For more details, check out the documentation on scaling CoreDNS.

What's next

You can configure CoreDNS to support many more use cases than kube-dns does by modifying the CoreDNS configuration ("Corefile"). For more information, see the documentation for the kubernetes CoreDNS plugin, or read the Custom DNS Entries for Kubernetes. in the CoreDNS blog.

42 - Using NodeLocal DNSCache in Kubernetes Clusters

FEATURE STATE: Kubernetes v1.18 [stable]

This page provides an overview of NodeLocal DNSCache feature in Kubernetes.

Before you begin

You need to have a Kubernetes cluster, and the kubectl command-line tool must be configured to communicate with your cluster. It is recommended to run this tutorial on a cluster with at least two nodes that are not acting as control plane hosts. If you do not already have a cluster, you can create one by using minikube or you can use one of these Kubernetes playgrounds:

To check the version, enter kubectl version.

Introduction

NodeLocal DNSCache improves Cluster DNS performance by running a DNS caching agent on cluster nodes as a DaemonSet. In today's architecture, Pods in 'ClusterFirst' DNS mode reach out to a kube-dns serviceIP for DNS queries. This is translated to a kube-dns/CoreDNS endpoint via iptables rules added by kube-proxy. With this new architecture, Pods will reach out to the DNS caching agent running on the same node, thereby avoiding iptables DNAT rules and connection tracking. The local caching agent will query kube-dns service for cache misses of cluster hostnames ("cluster.local" suffix by default).

Motivation

  • With the current DNS architecture, it is possible that Pods with the highest DNS QPS have to reach out to a different node, if there is no local kube-dns/CoreDNS instance. Having a local cache will help improve the latency in such scenarios.

  • Skipping iptables DNAT and connection tracking will help reduce conntrack races and avoid UDP DNS entries filling up conntrack table.

  • Connections from the local caching agent to kube-dns service can be upgraded to TCP. TCP conntrack entries will be removed on connection close in contrast with UDP entries that have to timeout (default nf_conntrack_udp_timeout is 30 seconds)

  • Upgrading DNS queries from UDP to TCP would reduce tail latency attributed to dropped UDP packets and DNS timeouts usually up to 30s (3 retries + 10s timeout). Since the nodelocal cache listens for UDP DNS queries, applications don't need to be changed.

  • Metrics & visibility into DNS requests at a node level.

  • Negative caching can be re-enabled, thereby reducing the number of queries for the kube-dns service.

Architecture Diagram

This is the path followed by DNS Queries after NodeLocal DNSCache is enabled:

NodeLocal DNSCache flow

Nodelocal DNSCache flow

This image shows how NodeLocal DNSCache handles DNS queries.

Configuration

This feature can be enabled using the following steps:

  • Prepare a manifest similar to the sample nodelocaldns.yaml and save it as nodelocaldns.yaml.

  • If using IPv6, the CoreDNS configuration file needs to enclose all the IPv6 addresses into square brackets if used in 'IP:Port' format. If you are using the sample manifest from the previous point, this will require you to modify the configuration line L70 like this: "health [__PILLAR__LOCAL__DNS__]:8080"

  • Substitute the variables in the manifest with the right values:

    kubedns=`kubectl get svc kube-dns -n kube-system -o jsonpath={.spec.clusterIP}`
domain=<cluster-domain>
localdns=<node-local-address>

    <cluster-domain> is "cluster.local" by default. <node-local-address> is the local listen IP address chosen for NodeLocal DNSCache.

    • If kube-proxy is running in IPTABLES mode:

      sed -i "s/__PILLAR__LOCAL__DNS__/$localdns/g; s/__PILLAR__DNS__DOMAIN__/$domain/g; s/__PILLAR__DNS__SERVER__/$kubedns/g" nodelocaldns.yaml

      __PILLAR__CLUSTER__DNS__ and __PILLAR__UPSTREAM__SERVERS__ will be populated by the node-local-dns pods. In this mode, the node-local-dns pods listen on both the kube-dns service IP as well as <node-local-address>, so pods can look up DNS records using either IP address.

    • If kube-proxy is running in IPVS mode:

      sed -i "s/__PILLAR__LOCAL__DNS__/$localdns/g; s/__PILLAR__DNS__DOMAIN__/$domain/g; s/,__PILLAR__DNS__SERVER__//g; s/__PILLAR__CLUSTER__DNS__/$kubedns/g" nodelocaldns.yaml

      In this mode, the node-local-dns pods listen only on <node-local-address>. The node-local-dns interface cannot bind the kube-dns cluster IP since the interface used for IPVS loadbalancing already uses this address. __PILLAR__UPSTREAM__SERVERS__ will be populated by the node-local-dns pods.

  • Run kubectl create -f nodelocaldns.yaml

  • If using kube-proxy in IPVS mode, --cluster-dns flag to kubelet needs to be modified to use <node-local-address> that NodeLocal DNSCache is listening on. Otherwise, there is no need to modify the value of the --cluster-dns flag, since NodeLocal DNSCache listens on both the kube-dns service IP as well as <node-local-address>.

Once enabled, the node-local-dns Pods will run in the kube-system namespace on each of the cluster nodes. This Pod runs CoreDNS in cache mode, so all CoreDNS metrics exposed by the different plugins will be available on a per-node basis.

You can disable this feature by removing the DaemonSet, using kubectl delete -f <manifest>. You should also revert any changes you made to the kubelet configuration.

StubDomains and Upstream server Configuration

StubDomains and upstream servers specified in the kube-dns ConfigMap in the kube-system namespace are automatically picked up by node-local-dns pods. The ConfigMap contents need to follow the format shown in the example. The node-local-dns ConfigMap can also be modified directly with the stubDomain configuration in the Corefile format. Some cloud providers might not allow modifying node-local-dns ConfigMap directly. In those cases, the kube-dns ConfigMap can be updated.

Setting memory limits

The node-local-dns Pods use memory for storing cache entries and processing queries. Since they do not watch Kubernetes objects, the cluster size or the number of Services / EndpointSlices do not directly affect memory usage. Memory usage is influenced by the DNS query pattern. From CoreDNS docs,

The default cache size is 10000 entries, which uses about 30 MB when completely filled.

This would be the memory usage for each server block (if the cache gets completely filled). Memory usage can be reduced by specifying smaller cache sizes.

The number of concurrent queries is linked to the memory demand, because each extra goroutine used for handling a query requires an amount of memory. You can set an upper limit using the max_concurrent option in the forward plugin.

If a node-local-dns Pod attempts to use more memory than is available (because of total system resources, or because of a configured resource limit), the operating system may shut down that pod's container. If this happens, the container that is terminated (“OOMKilled”) does not clean up the custom packet filtering rules that it previously added during startup. The node-local-dns container should get restarted (since managed as part of a DaemonSet), but this will lead to a brief DNS downtime each time that the container fails: the packet filtering rules direct DNS queries to a local Pod that is unhealthy.

You can determine a suitable memory limit by running node-local-dns pods without a limit and measuring the peak usage. You can also set up and use a VerticalPodAutoscaler in recommender mode, and then check its recommendations.

43 - Using sysctls in a Kubernetes Cluster

FEATURE STATE: Kubernetes v1.21 [stable]

This document describes how to configure and use kernel parameters within a Kubernetes cluster using the sysctl interface.

Before you begin

You need to have a Kubernetes cluster, and the kubectl command-line tool must be configured to communicate with your cluster. It is recommended to run this tutorial on a cluster with at least two nodes that are not acting as control plane hosts. If you do not already have a cluster, you can create one by using minikube or you can use one of these Kubernetes playgrounds:

For some steps, you also need to be able to reconfigure the command line options for the kubelets running on your cluster.

Listing all Sysctl Parameters

In Linux, the sysctl interface allows an administrator to modify kernel parameters at runtime. Parameters are available via the /proc/sys/ virtual process file system. The parameters cover various subsystems such as:

  • kernel (common prefix: kernel.)
  • networking (common prefix: net.)
  • virtual memory (common prefix: vm.)
  • MDADM (common prefix: dev.)
  • More subsystems are described in Kernel docs.

To get a list of all parameters, you can run

sudo sysctl -a

Safe and Unsafe Sysctls

Kubernetes classes sysctls as either safe or unsafe. In addition to proper namespacing, a safe sysctl must be properly isolated between pods on the same node. This means that setting a safe sysctl for one pod

  • must not have any influence on any other pod on the node
  • must not allow to harm the node's health
  • must not allow to gain CPU or memory resources outside of the resource limits of a pod.

By far, most of the namespaced sysctls are not necessarily considered safe. The following sysctls are supported in the safe set:

  • kernel.shm_rmid_forced;
  • net.ipv4.ip_local_port_range;
  • net.ipv4.tcp_syncookies;
  • net.ipv4.ping_group_range (since Kubernetes 1.18);
  • net.ipv4.ip_unprivileged_port_start (since Kubernetes 1.22);
  • net.ipv4.ip_local_reserved_ports (since Kubernetes 1.27, needs kernel 3.16+);
  • net.ipv4.tcp_keepalive_time (since Kubernetes 1.29, needs kernel 4.5+);
  • net.ipv4.tcp_fin_timeout (since Kubernetes 1.29, needs kernel 4.6+);
  • net.ipv4.tcp_keepalive_intvl (since Kubernetes 1.29, needs kernel 4.5+);
  • net.ipv4.tcp_keepalive_probes (since Kubernetes 1.29, needs kernel 4.5+).
  • net.ipv4.tcp_rmem (since Kubernetes 1.32, needs kernel 4.15+).
  • net.ipv4.tcp_wmem (since Kubernetes 1.32, needs kernel 4.15+).

This list will be extended in future Kubernetes versions when the kubelet supports better isolation mechanisms.

Enabling Unsafe Sysctls

All safe sysctls are enabled by default.

All unsafe sysctls are disabled by default and must be allowed manually by the cluster admin on a per-node basis. Pods with disabled unsafe sysctls will be scheduled, but will fail to launch.

With the warning above in mind, the cluster admin can allow certain unsafe sysctls for very special situations such as high-performance or real-time application tuning. Unsafe sysctls are enabled on a node-by-node basis with a flag of the kubelet; for example:

kubelet --allowed-unsafe-sysctls \
  'kernel.msg*,net.core.somaxconn' ...

For Minikube, this can be done via the extra-config flag:

minikube start --extra-config="kubelet.allowed-unsafe-sysctls=kernel.msg*,net.core.somaxconn"...

Only namespaced sysctls can be enabled this way.

Setting Sysctls for a Pod

A number of sysctls are namespaced in today's Linux kernels. This means that they can be set independently for each pod on a node. Only namespaced sysctls are configurable via the pod securityContext within Kubernetes.

The following sysctls are known to be namespaced. This list could change in future versions of the Linux kernel.

  • kernel.shm*,
  • kernel.msg*,
  • kernel.sem,
  • fs.mqueue.*,
  • Those net.* that can be set in container networking namespace. However, there are exceptions (e.g., net.netfilter.nf_conntrack_max and net.netfilter.nf_conntrack_expect_max can be set in container networking namespace but are unnamespaced before Linux 5.12.2).

Sysctls with no namespace are called node-level sysctls. If you need to set them, you must manually configure them on each node's operating system, or by using a DaemonSet with privileged containers.

Use the pod securityContext to configure namespaced sysctls. The securityContext applies to all containers in the same pod.

This example uses the pod securityContext to set a safe sysctl kernel.shm_rmid_forced and two unsafe sysctls net.core.somaxconn and kernel.msgmax. There is no distinction between safe and unsafe sysctls in the specification.

apiVersion: v1
kind: Pod
metadata:
  name: sysctl-example
spec:
  securityContext:
    sysctls:
    - name: kernel.shm_rmid_forced
      value: "0"
    - name: net.core.somaxconn
      value: "1024"
    - name: kernel.msgmax
      value: "65536"
  ...

It is good practice to consider nodes with special sysctl settings as tainted within a cluster, and only schedule pods onto them which need those sysctl settings. It is suggested to use the Kubernetes taints and toleration feature to implement this.

A pod with the unsafe sysctls will fail to launch on any node which has not enabled those two unsafe sysctls explicitly. As with node-level sysctls it is recommended to use taints and toleration feature or taints on nodes to schedule those pods onto the right nodes.

44 - Utilizing the NUMA-aware Memory Manager

FEATURE STATE: Kubernetes v1.32 [stable] (enabled by default: true)

The Kubernetes Memory Manager enables the feature of guaranteed memory (and hugepages) allocation for pods in the Guaranteed QoS class.

The Memory Manager employs hint generation protocol to yield the most suitable NUMA affinity for a pod. The Memory Manager feeds the central manager (Topology Manager) with these affinity hints. Based on both the hints and Topology Manager policy, the pod is rejected or admitted to the node.

Moreover, the Memory Manager ensures that the memory which a pod requests is allocated from a minimum number of NUMA nodes.

The Memory Manager is only pertinent to Linux based hosts.

Before you begin

You need to have a Kubernetes cluster, and the kubectl command-line tool must be configured to communicate with your cluster. It is recommended to run this tutorial on a cluster with at least two nodes that are not acting as control plane hosts. If you do not already have a cluster, you can create one by using minikube or you can use one of these Kubernetes playgrounds:

Your Kubernetes server must be version v1.32. To check the version, enter kubectl version.

To align memory resources with other requested resources in a Pod spec:

Starting from v1.22, the Memory Manager is enabled by default through MemoryManager feature gate.

Preceding v1.22, the kubelet must be started with the following flag:

--feature-gates=MemoryManager=true

in order to enable the Memory Manager feature.

How does the Memory Manager Operate?

The Memory Manager currently offers the guaranteed memory (and hugepages) allocation for Pods in Guaranteed QoS class. To immediately put the Memory Manager into operation follow the guidelines in the section Memory Manager configuration, and subsequently, prepare and deploy a Guaranteed pod as illustrated in the section Placing a Pod in the Guaranteed QoS class.

The Memory Manager is a Hint Provider, and it provides topology hints for the Topology Manager which then aligns the requested resources according to these topology hints. On Linux, it also enforces cgroups (i.e. cpuset.mems) for pods. The complete flow diagram concerning pod admission and deployment process is illustrated in Memory Manager KEP: Design Overview and below:

Memory Manager in the pod admission and deployment process

During this process, the Memory Manager updates its internal counters stored in Node Map and Memory Maps to manage guaranteed memory allocation.

The Memory Manager updates the Node Map during the startup and runtime as follows.

Startup

This occurs once a node administrator employs --reserved-memory (section Reserved memory flag). In this case, the Node Map becomes updated to reflect this reservation as illustrated in Memory Manager KEP: Memory Maps at start-up (with examples).

The administrator must provide --reserved-memory flag when Static policy is configured.

Runtime

Reference Memory Manager KEP: Memory Maps at runtime (with examples) illustrates how a successful pod deployment affects the Node Map, and it also relates to how potential Out-of-Memory (OOM) situations are handled further by Kubernetes or operating system.

Important topic in the context of Memory Manager operation is the management of NUMA groups. Each time pod's memory request is in excess of single NUMA node capacity, the Memory Manager attempts to create a group that comprises several NUMA nodes and features extend memory capacity. The problem has been solved as elaborated in Memory Manager KEP: How to enable the guaranteed memory allocation over many NUMA nodes?. Also, reference Memory Manager KEP: Simulation - how the Memory Manager works? (by examples) illustrates how the management of groups occurs.

Windows Support

FEATURE STATE: Kubernetes v1.32 [alpha] (enabled by default: false)

Windows support can be enabled via the WindowsCPUAndMemoryAffinity feature gate and it requires support in the container runtime. Only the BestEffort Policy is supported on Windows.

Memory Manager configuration

Other Managers should be first pre-configured. Next, the Memory Manager feature should be enabled and be run with Static policy (section Static policy). Optionally, some amount of memory can be reserved for system or kubelet processes to increase node stability (section Reserved memory flag).

Policies

Memory Manager supports two policies. You can select a policy via a kubelet flag --memory-manager-policy:

  • None (default)
  • Static (Linux only)
  • BestEffort (Windows Only)

None policy

This is the default policy and does not affect the memory allocation in any way. It acts the same as if the Memory Manager is not present at all.

The None policy returns default topology hint. This special hint denotes that Hint Provider (Memory Manager in this case) has no preference for NUMA affinity with any resource.

Static policy

In the case of the Guaranteed pod, the Static Memory Manager policy returns topology hints relating to the set of NUMA nodes where the memory can be guaranteed, and reserves the memory through updating the internal NodeMap object.

In the case of the BestEffort or Burstable pod, the Static Memory Manager policy sends back the default topology hint as there is no request for the guaranteed memory, and does not reserve the memory in the internal NodeMap object.

This policy is only supported on Linux.

BestEffort policy

FEATURE STATE: Kubernetes v1.32 [alpha] (enabled by default: false)

This policy is only supported on Windows.

On Windows, NUMA node assignment works differently than Linux. There is no mechanism to ensure that Memory access only comes from a specific NUMA node. Instead the Windows scheduler will select the most optimal NUMA node based on the CPU(s) assignments. It is possible that Windows might use other NUMA nodes if deemed optimal by the Windows scheduler.

The policy does track the amount of memory available and requested through the internal NodeMap. The memory manager will make a best effort at ensuring that enough memory is available on a NUMA node before making the assignment.
This means that in most cases memory assignment should function as expected.

Reserved memory flag

The Node Allocatable mechanism is commonly used by node administrators to reserve K8S node system resources for the kubelet or operating system processes in order to enhance the node stability. A dedicated set of flags can be used for this purpose to set the total amount of reserved memory for a node. This pre-configured value is subsequently utilized to calculate the real amount of node's "allocatable" memory available to pods.

The Kubernetes scheduler incorporates "allocatable" to optimise pod scheduling process. The foregoing flags include --kube-reserved, --system-reserved and --eviction-threshold. The sum of their values will account for the total amount of reserved memory.

A new --reserved-memory flag was added to Memory Manager to allow for this total reserved memory to be split (by a node administrator) and accordingly reserved across many NUMA nodes.

The flag specifies a comma-separated list of memory reservations of different memory types per NUMA node. Memory reservations across multiple NUMA nodes can be specified using semicolon as separator. This parameter is only useful in the context of the Memory Manager feature. The Memory Manager will not use this reserved memory for the allocation of container workloads.

For example, if you have a NUMA node "NUMA0" with 10Gi of memory available, and the --reserved-memory was specified to reserve 1Gi of memory at "NUMA0", the Memory Manager assumes that only 9Gi is available for containers.

You can omit this parameter, however, you should be aware that the quantity of reserved memory from all NUMA nodes should be equal to the quantity of memory specified by the Node Allocatable feature. If at least one node allocatable parameter is non-zero, you will need to specify --reserved-memory for at least one NUMA node. In fact, eviction-hard threshold value is equal to 100Mi by default, so if Static policy is used, --reserved-memory is obligatory.

Also, avoid the following configurations:

  1. duplicates, i.e. the same NUMA node or memory type, but with a different value;
  2. setting zero limit for any of memory types;
  3. NUMA node IDs that do not exist in the machine hardware;
  4. memory type names different than memory or hugepages-<size> (hugepages of particular <size> should also exist).

Syntax:

--reserved-memory N:memory-type1=value1,memory-type2=value2,...

  • N (integer) - NUMA node index, e.g. 0
  • memory-type (string) - represents memory type:
    • memory - conventional memory
    • hugepages-2Mi or hugepages-1Gi - hugepages
  • value (string) - the quantity of reserved memory, e.g. 1Gi

Example usage:

--reserved-memory 0:memory=1Gi,hugepages-1Gi=2Gi

or

--reserved-memory 0:memory=1Gi --reserved-memory 1:memory=2Gi

or

--reserved-memory '0:memory=1Gi;1:memory=2Gi'

When you specify values for --reserved-memory flag, you must comply with the setting that you prior provided via Node Allocatable Feature flags. That is, the following rule must be obeyed for each memory type:

sum(reserved-memory(i)) = kube-reserved + system-reserved + eviction-threshold,

where i is an index of a NUMA node.

If you do not follow the formula above, the Memory Manager will show an error on startup.

In other words, the example above illustrates that for the conventional memory (type=memory), we reserve 3Gi in total, i.e.:

sum(reserved-memory(i)) = reserved-memory(0) + reserved-memory(1) = 1Gi + 2Gi = 3Gi

An example of kubelet command-line arguments relevant to the node Allocatable configuration:

  • --kube-reserved=cpu=500m,memory=50Mi
  • --system-reserved=cpu=123m,memory=333Mi
  • --eviction-hard=memory.available<500Mi

Here is an example of a correct configuration:

--kube-reserved=cpu=4,memory=4Gi
--system-reserved=cpu=1,memory=1Gi
--memory-manager-policy=Static
--reserved-memory '0:memory=3Gi;1:memory=2148Mi'

Prior to Kubernetes 1.32, you also need to add

--feature-gates=MemoryManager=true

Let us validate the configuration above:

  1. kube-reserved + system-reserved + eviction-hard(default) = reserved-memory(0) + reserved-memory(1)
  2. 4GiB + 1GiB + 100MiB = 3GiB + 2148MiB
  3. 5120MiB + 100MiB = 3072MiB + 2148MiB
  4. 5220MiB = 5220MiB (which is correct)

Placing a Pod in the Guaranteed QoS class

If the selected policy is anything other than None, the Memory Manager identifies pods that are in the Guaranteed QoS class. The Memory Manager provides specific topology hints to the Topology Manager for each Guaranteed pod. For pods in a QoS class other than Guaranteed, the Memory Manager provides default topology hints to the Topology Manager.

The following excerpts from pod manifests assign a pod to the Guaranteed QoS class.

Pod with integer CPU(s) runs in the Guaranteed QoS class, when requests are equal to limits:

spec:
  containers:
  - name: nginx
    image: nginx
    resources:
      limits:
        memory: "200Mi"
        cpu: "2"
        example.com/device: "1"
      requests:
        memory: "200Mi"
        cpu: "2"
        example.com/device: "1"

Also, a pod sharing CPU(s) runs in the Guaranteed QoS class, when requests are equal to limits.

spec:
  containers:
  - name: nginx
    image: nginx
    resources:
      limits:
        memory: "200Mi"
        cpu: "300m"
        example.com/device: "1"
      requests:
        memory: "200Mi"
        cpu: "300m"
        example.com/device: "1"

Notice that both CPU and memory requests must be specified for a Pod to lend it to Guaranteed QoS class.

Troubleshooting

The following means can be used to troubleshoot the reason why a pod could not be deployed or became rejected at a node:

  • pod status - indicates topology affinity errors
  • system logs - include valuable information for debugging, e.g., about generated hints
  • state file - the dump of internal state of the Memory Manager (includes Node Map and Memory Maps)
  • starting from v1.22, the device plugin resource API can be used to retrieve information about the memory reserved for containers

Pod status (TopologyAffinityError)

This error typically occurs in the following situations:

  • a node has not enough resources available to satisfy the pod's request
  • the pod's request is rejected due to particular Topology Manager policy constraints

The error appears in the status of a pod:

kubectl get pods

NAME         READY   STATUS                  RESTARTS   AGE
guaranteed   0/1     TopologyAffinityError   0          113s

Use kubectl describe pod <id> or kubectl get events to obtain detailed error message:

Warning  TopologyAffinityError  10m   kubelet, dell8  Resources cannot be allocated with Topology locality

System logs

Search system logs with respect to a particular pod.

The set of hints that Memory Manager generated for the pod can be found in the logs. Also, the set of hints generated by CPU Manager should be present in the logs.

Topology Manager merges these hints to calculate a single best hint. The best hint should be also present in the logs.

The best hint indicates where to allocate all the resources. Topology Manager tests this hint against its current policy, and based on the verdict, it either admits the pod to the node or rejects it.

Also, search the logs for occurrences associated with the Memory Manager, e.g. to find out information about cgroups and cpuset.mems updates.

Examine the memory manager state on a node

Let us first deploy a sample Guaranteed pod whose specification is as follows:

apiVersion: v1
kind: Pod
metadata:
  name: guaranteed
spec:
  containers:
  - name: guaranteed
    image: consumer
    imagePullPolicy: Never
    resources:
      limits:
        cpu: "2"
        memory: 150Gi
      requests:
        cpu: "2"
        memory: 150Gi
    command: ["sleep","infinity"]

Next, let us log into the node where it was deployed and examine the state file in /var/lib/kubelet/memory_manager_state:

{
   "policyName":"Static",
   "machineState":{
      "0":{
         "numberOfAssignments":1,
         "memoryMap":{
            "hugepages-1Gi":{
               "total":0,
               "systemReserved":0,
               "allocatable":0,
               "reserved":0,
               "free":0
            },
            "memory":{
               "total":134987354112,
               "systemReserved":3221225472,
               "allocatable":131766128640,
               "reserved":131766128640,
               "free":0
            }
         },
         "nodes":[
            0,
            1
         ]
      },
      "1":{
         "numberOfAssignments":1,
         "memoryMap":{
            "hugepages-1Gi":{
               "total":0,
               "systemReserved":0,
               "allocatable":0,
               "reserved":0,
               "free":0
            },
            "memory":{
               "total":135286722560,
               "systemReserved":2252341248,
               "allocatable":133034381312,
               "reserved":29295144960,
               "free":103739236352
            }
         },
         "nodes":[
            0,
            1
         ]
      }
   },
   "entries":{
      "fa9bdd38-6df9-4cf9-aa67-8c4814da37a8":{
         "guaranteed":[
            {
               "numaAffinity":[
                  0,
                  1
               ],
               "type":"memory",
               "size":161061273600
            }
         ]
      }
   },
   "checksum":4142013182
}

It can be deduced from the state file that the pod was pinned to both NUMA nodes, i.e.:

"numaAffinity":[
   0,
   1
],

Pinned term means that pod's memory consumption is constrained (through cgroups configuration) to these NUMA nodes.

This automatically implies that Memory Manager instantiated a new group that comprises these two NUMA nodes, i.e. 0 and 1 indexed NUMA nodes.

Notice that the management of groups is handled in a relatively complex manner, and further elaboration is provided in Memory Manager KEP in this and this sections.

In order to analyse memory resources available in a group,the corresponding entries from NUMA nodes belonging to the group must be added up.

For example, the total amount of free "conventional" memory in the group can be computed by adding up the free memory available at every NUMA node in the group, i.e., in the "memory" section of NUMA node 0 ("free":0) and NUMA node 1 ("free":103739236352). So, the total amount of free "conventional" memory in this group is equal to 0 + 103739236352 bytes.

The line "systemReserved":3221225472 indicates that the administrator of this node reserved 3221225472 bytes (i.e. 3Gi) to serve kubelet and system processes at NUMA node 0, by using --reserved-memory flag.

Device plugin resource API

The kubelet provides a PodResourceLister gRPC service to enable discovery of resources and associated metadata. By using its List gRPC endpoint, information about reserved memory for each container can be retrieved, which is contained in protobuf ContainerMemory message. This information can be retrieved solely for pods in Guaranteed QoS class.

What's next

45 - Verify Signed Kubernetes Artifacts

FEATURE STATE: Kubernetes v1.26 [beta]

Before you begin

You will need to have the following tools installed:

Verifying binary signatures

The Kubernetes release process signs all binary artifacts (tarballs, SPDX files, standalone binaries) by using cosign's keyless signing. To verify a particular binary, retrieve it together with its signature and certificate:

URL=https://dl.k8s.io/release/v1.32.0/bin/linux/amd64
BINARY=kubectl

FILES=(
    "$BINARY"
    "$BINARY.sig"
    "$BINARY.cert"
)

for FILE in "${FILES[@]}"; do
    curl -sSfL --retry 3 --retry-delay 3 "$URL/$FILE" -o "$FILE"
done

Then verify the blob by using cosign verify-blob:

cosign verify-blob "$BINARY" \
  --signature "$BINARY".sig \
  --certificate "$BINARY".cert \
  --certificate-identity krel-staging@k8s-releng-prod.iam.gserviceaccount.com \
  --certificate-oidc-issuer https://accounts.google.com

Verifying image signatures

For a complete list of images that are signed please refer to Releases.

Pick one image from this list and verify its signature using the cosign verify command:

cosign verify registry.k8s.io/kube-apiserver-amd64:v1.32.0 \
  --certificate-identity krel-trust@k8s-releng-prod.iam.gserviceaccount.com \
  --certificate-oidc-issuer https://accounts.google.com \
  | jq .

Verifying images for all control plane components

To verify all signed control plane images for the latest stable version (v1.32.0), please run the following commands:

curl -Ls "https://sbom.k8s.io/$(curl -Ls https://dl.k8s.io/release/stable.txt)/release" \
  | grep "SPDXID: SPDXRef-Package-registry.k8s.io" \
  | grep -v sha256 | cut -d- -f3- | sed 's/-/\//' | sed 's/-v1/:v1/' \
  | sort > images.txt
input=images.txt
while IFS= read -r image
do
  cosign verify "$image" \
    --certificate-identity krel-trust@k8s-releng-prod.iam.gserviceaccount.com \
    --certificate-oidc-issuer https://accounts.google.com \
    | jq .
done < "$input"

Once you have verified an image, you can specify the image by its digest in your Pod manifests as per this example:

registry-url/image-name@sha256:45b23dee08af5e43a7fea6c4cf9c25ccf269ee113168c19722f87876677c5cb2

For more information, please refer to the Image Pull Policy section.

Verifying Image Signatures with Admission Controller

For non-control plane images (for example conformance image), signatures can also be verified at deploy time using sigstore policy-controller admission controller.

Here are some helpful resources to get started with policy-controller:

Verify the Software Bill Of Materials

You can verify the Kubernetes Software Bill of Materials (SBOM) by using the sigstore certificate and signature, or the corresponding SHA files:

# Retrieve the latest available Kubernetes release version
VERSION=$(curl -Ls https://dl.k8s.io/release/stable.txt)

# Verify the SHA512 sum
curl -Ls "https://sbom.k8s.io/$VERSION/release" -o "$VERSION.spdx"
echo "$(curl -Ls "https://sbom.k8s.io/$VERSION/release.sha512") $VERSION.spdx" | sha512sum --check

# Verify the SHA256 sum
echo "$(curl -Ls "https://sbom.k8s.io/$VERSION/release.sha256") $VERSION.spdx" | sha256sum --check

# Retrieve sigstore signature and certificate
curl -Ls "https://sbom.k8s.io/$VERSION/release.sig" -o "$VERSION.spdx.sig"
curl -Ls "https://sbom.k8s.io/$VERSION/release.cert" -o "$VERSION.spdx.cert"

# Verify the sigstore signature
cosign verify-blob \
    --certificate "$VERSION.spdx.cert" \
    --signature "$VERSION.spdx.sig" \
    --certificate-identity krel-staging@k8s-releng-prod.iam.gserviceaccount.com \
    --certificate-oidc-issuer https://accounts.google.com \
    "$VERSION.spdx"