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

Return to the regular view of this page.

Cluster Administration

Lower-level detail relevant to creating or administering a Kubernetes cluster.

The cluster administration overview is for anyone creating or administering a Kubernetes cluster. It assumes some familiarity with core Kubernetes concepts.

Planning a cluster

See the guides in Setup for examples of how to plan, set up, and configure Kubernetes clusters. The solutions listed in this article are called distros.

Before choosing a guide, here are some considerations:

  • Do you want to try out Kubernetes on your computer, or do you want to build a high-availability, multi-node cluster? Choose distros best suited for your needs.
  • Will you be using a hosted Kubernetes cluster, such as Google Kubernetes Engine, or hosting your own cluster?
  • Will your cluster be on-premises, or in the cloud (IaaS)? Kubernetes does not directly support hybrid clusters. Instead, you can set up multiple clusters.
  • If you are configuring Kubernetes on-premises, consider which networking model fits best.
  • Will you be running Kubernetes on "bare metal" hardware or on virtual machines (VMs)?
  • Do you want to run a cluster, or do you expect to do active development of Kubernetes project code? If the latter, choose an actively-developed distro. Some distros only use binary releases, but offer a greater variety of choices.
  • Familiarize yourself with the components needed to run a cluster.

Managing a cluster

Securing a cluster

Securing the kubelet

Optional Cluster Services

1 - Node Shutdowns

In a Kubernetes cluster, a node can be shutdown in a planned graceful way or unexpectedly because of reasons such as a power outage or something else external. A node shutdown could lead to workload failure if the node is not drained before the shutdown. A node shutdown can be either graceful or non-graceful.

Graceful node shutdown

FEATURE STATE: Kubernetes v1.21 [beta]

The kubelet attempts to detect node system shutdown and terminates pods running on the node.

Kubelet ensures that pods follow the normal pod termination process during the node shutdown. During node shutdown, the kubelet does not accept new Pods (even if those Pods are already bound to the node).

The Graceful node shutdown feature depends on systemd since it takes advantage of systemd inhibitor locks to delay the node shutdown with a given duration.

Graceful node shutdown is controlled with the GracefulNodeShutdown feature gate which is enabled by default in 1.21.

Note that by default, both configuration options described below, shutdownGracePeriod and shutdownGracePeriodCriticalPods are set to zero, thus not activating the graceful node shutdown functionality. To activate the feature, the two kubelet config settings should be configured appropriately and set to non-zero values.

Once systemd detects or notifies node shutdown, the kubelet sets a NotReady condition on the Node, with the reason set to "node is shutting down". The kube-scheduler honors this condition and does not schedule any Pods onto the affected node; other third-party schedulers are expected to follow the same logic. This means that new Pods won't be scheduled onto that node and therefore none will start.

The kubelet also rejects Pods during the PodAdmission phase if an ongoing node shutdown has been detected, so that even Pods with a toleration for node.kubernetes.io/not-ready:NoSchedule do not start there.

At the same time when kubelet is setting that condition on its Node via the API, the kubelet also begins terminating any Pods that are running locally.

During a graceful shutdown, kubelet terminates pods in two phases:

  1. Terminate regular pods running on the node.
  2. Terminate critical pods running on the node.

Graceful node shutdown feature is configured with two KubeletConfiguration options:

  • shutdownGracePeriod:
    • Specifies the total duration that the node should delay the shutdown by. This is the total grace period for pod termination for both regular and critical pods.
  • shutdownGracePeriodCriticalPods:
    • Specifies the duration used to terminate critical pods during a node shutdown. This value should be less than shutdownGracePeriod.

For example, if shutdownGracePeriod=30s, and shutdownGracePeriodCriticalPods=10s, kubelet will delay the node shutdown by 30 seconds. During the shutdown, the first 20 (30-10) seconds would be reserved for gracefully terminating normal pods, and the last 10 seconds would be reserved for terminating critical pods.

Pod Priority based graceful node shutdown

FEATURE STATE: Kubernetes v1.24 [beta]

To provide more flexibility during graceful node shutdown around the ordering of pods during shutdown, graceful node shutdown honors the PriorityClass for Pods, provided that you enabled this feature in your cluster. The feature allows cluster administers to explicitly define the ordering of pods during graceful node shutdown based on priority classes.

The Graceful Node Shutdown feature, as described above, shuts down pods in two phases, non-critical pods, followed by critical pods. If additional flexibility is needed to explicitly define the ordering of pods during shutdown in a more granular way, pod priority based graceful shutdown can be used.

When graceful node shutdown honors pod priorities, this makes it possible to do graceful node shutdown in multiple phases, each phase shutting down a particular priority class of pods. The kubelet can be configured with the exact phases and shutdown time per phase.

Assuming the following custom pod priority classes in a cluster,

Pod priority class name Pod priority class value
custom-class-a 100000
custom-class-b 10000
custom-class-c 1000
regular/unset 0

Within the kubelet configuration the settings for shutdownGracePeriodByPodPriority could look like:

Pod priority class value Shutdown period
100000 10 seconds
10000 180 seconds
1000 120 seconds
0 60 seconds

The corresponding kubelet config YAML configuration would be:

shutdownGracePeriodByPodPriority:
  - priority: 100000
    shutdownGracePeriodSeconds: 10
  - priority: 10000
    shutdownGracePeriodSeconds: 180
  - priority: 1000
    shutdownGracePeriodSeconds: 120
  - priority: 0
    shutdownGracePeriodSeconds: 60

The above table implies that any pod with priority value >= 100000 will get just 10 seconds to stop, any pod with value >= 10000 and < 100000 will get 180 seconds to stop, any pod with value >= 1000 and < 10000 will get 120 seconds to stop. Finally, all other pods will get 60 seconds to stop.

One doesn't have to specify values corresponding to all of the classes. For example, you could instead use these settings:

Pod priority class value Shutdown period
100000 300 seconds
1000 120 seconds
0 60 seconds

In the above case, the pods with custom-class-b will go into the same bucket as custom-class-c for shutdown.

If there are no pods in a particular range, then the kubelet does not wait for pods in that priority range. Instead, the kubelet immediately skips to the next priority class value range.

If this feature is enabled and no configuration is provided, then no ordering action will be taken.

Using this feature requires enabling the GracefulNodeShutdownBasedOnPodPriority feature gate, and setting ShutdownGracePeriodByPodPriority in the kubelet config to the desired configuration containing the pod priority class values and their respective shutdown periods.

Metrics graceful_shutdown_start_time_seconds and graceful_shutdown_end_time_seconds are emitted under the kubelet subsystem to monitor node shutdowns.

Non-graceful node shutdown handling

FEATURE STATE: Kubernetes v1.28 [stable]

A node shutdown action may not be detected by kubelet's Node Shutdown Manager, either because the command does not trigger the inhibitor locks mechanism used by kubelet or because of a user error, i.e., the ShutdownGracePeriod and ShutdownGracePeriodCriticalPods are not configured properly. Please refer to above section Graceful Node Shutdown for more details.

When a node is shutdown but not detected by kubelet's Node Shutdown Manager, the pods that are part of a StatefulSet will be stuck in terminating status on the shutdown node and cannot move to a new running node. This is because kubelet on the shutdown node is not available to delete the pods so the StatefulSet cannot create a new pod with the same name. If there are volumes used by the pods, the VolumeAttachments will not be deleted from the original shutdown node so the volumes used by these pods cannot be attached to a new running node. As a result, the application running on the StatefulSet cannot function properly. If the original shutdown node comes up, the pods will be deleted by kubelet and new pods will be created on a different running node. If the original shutdown node does not come up, these pods will be stuck in terminating status on the shutdown node forever.

To mitigate the above situation, a user can manually add the taint node.kubernetes.io/out-of-service with either NoExecute or NoSchedule effect to a Node marking it out-of-service. If the NodeOutOfServiceVolumeDetachfeature gate is enabled on kube-controller-manager, and a Node is marked out-of-service with this taint, the pods on the node will be forcefully deleted if there are no matching tolerations on it and volume detach operations for the pods terminating on the node will happen immediately. This allows the Pods on the out-of-service node to recover quickly on a different node.

During a non-graceful shutdown, Pods are terminated in the two phases:

  1. Force delete the Pods that do not have matching out-of-service tolerations.
  2. Immediately perform detach volume operation for such pods.

Forced storage detach on timeout

In any situation where a pod deletion has not succeeded for 6 minutes, kubernetes will force detach volumes being unmounted if the node is unhealthy at that instant. Any workload still running on the node that uses a force-detached volume will cause a violation of the CSI specification, which states that ControllerUnpublishVolume "must be called after all NodeUnstageVolume and NodeUnpublishVolume on the volume are called and succeed". In such circumstances, volumes on the node in question might encounter data corruption.

The forced storage detach behaviour is optional; users might opt to use the "Non-graceful node shutdown" feature instead.

Force storage detach on timeout can be disabled by setting the disable-force-detach-on-timeout config field in kube-controller-manager. Disabling the force detach on timeout feature means that a volume that is hosted on a node that is unhealthy for more than 6 minutes will not have its associated VolumeAttachment deleted.

After this setting has been applied, unhealthy pods still attached to a volumes must be recovered via the Non-Graceful Node Shutdown procedure mentioned above.

What's next

Learn more about the following:

2 - Certificates

To learn how to generate certificates for your cluster, see Certificates.

3 - Cluster Networking

Networking is a central part of Kubernetes, but it can be challenging to understand exactly how it is expected to work. There are 4 distinct networking problems to address:

  1. Highly-coupled container-to-container communications: this is solved by Pods and localhost communications.
  2. Pod-to-Pod communications: this is the primary focus of this document.
  3. Pod-to-Service communications: this is covered by Services.
  4. External-to-Service communications: this is also covered by Services.

Kubernetes is all about sharing machines among applications. Typically, sharing machines requires ensuring that two applications do not try to use the same ports. Coordinating ports across multiple developers is very difficult to do at scale and exposes users to cluster-level issues outside of their control.

Dynamic port allocation brings a lot of complications to the system - every application has to take ports as flags, the API servers have to know how to insert dynamic port numbers into configuration blocks, services have to know how to find each other, etc. Rather than deal with this, Kubernetes takes a different approach.

To learn about the Kubernetes networking model, see here.

Kubernetes IP address ranges

Kubernetes clusters require to allocate non-overlapping IP addresses for Pods, Services and Nodes, from a range of available addresses configured in the following components:

  • The network plugin is configured to assign IP addresses to Pods.
  • The kube-apiserver is configured to assign IP addresses to Services.
  • The kubelet or the cloud-controller-manager is configured to assign IP addresses to Nodes.
A figure illustrating the different network ranges in a kubernetes cluster

Cluster networking types

Kubernetes clusters, attending to the IP families configured, can be categorized into:

  • IPv4 only: The network plugin, kube-apiserver and kubelet/cloud-controller-manager are configured to assign only IPv4 addresses.
  • IPv6 only: The network plugin, kube-apiserver and kubelet/cloud-controller-manager are configured to assign only IPv6 addresses.
  • IPv4/IPv6 or IPv6/IPv4 dual-stack:
    • The network plugin is configured to assign IPv4 and IPv6 addresses.
    • The kube-apiserver is configured to assign IPv4 and IPv6 addresses.
    • The kubelet or cloud-controller-manager is configured to assign IPv4 and IPv6 address.
    • All components must agree on the configured primary IP family.

Kubernetes clusters only consider the IP families present on the Pods, Services and Nodes objects, independently of the existing IPs of the represented objects. Per example, a server or a pod can have multiple IP addresses on its interfaces, but only the IP addresses in node.status.addresses or pod.status.ips are considered for implementing the Kubernetes network model and defining the type of the cluster.

How to implement the Kubernetes network model

The network model is implemented by the container runtime on each node. The most common container runtimes use Container Network Interface (CNI) plugins to manage their network and security capabilities. Many different CNI plugins exist from many different vendors. Some of these provide only basic features of adding and removing network interfaces, while others provide more sophisticated solutions, such as integration with other container orchestration systems, running multiple CNI plugins, advanced IPAM features etc.

See this page for a non-exhaustive list of networking addons supported by Kubernetes.

What's next

The early design of the networking model and its rationale are described in more detail in the networking design document. For future plans and some on-going efforts that aim to improve Kubernetes networking, please refer to the SIG-Network KEPs.

4 - Logging Architecture

Application logs can help you understand what is happening inside your application. The logs are particularly useful for debugging problems and monitoring cluster activity. Most modern applications have some kind of logging mechanism. Likewise, container engines are designed to support logging. The easiest and most adopted logging method for containerized applications is writing to standard output and standard error streams.

However, the native functionality provided by a container engine or runtime is usually not enough for a complete logging solution.

For example, you may want to access your application's logs if a container crashes, a pod gets evicted, or a node dies.

In a cluster, logs should have a separate storage and lifecycle independent of nodes, pods, or containers. This concept is called cluster-level logging.

Cluster-level logging architectures require a separate backend to store, analyze, and query logs. Kubernetes does not provide a native storage solution for log data. Instead, there are many logging solutions that integrate with Kubernetes. The following sections describe how to handle and store logs on nodes.

Pod and container logs

Kubernetes captures logs from each container in a running Pod.

This example uses a manifest for a Pod with a container that writes text to the standard output stream, once per second.

apiVersion: v1
kind: Pod
metadata:
  name: counter
spec:
  containers:
  - name: count
    image: busybox:1.28
    args: [/bin/sh, -c,
            'i=0; while true; do echo "$i: $(date)"; i=$((i+1)); sleep 1; done']

To run this pod, use the following command:

kubectl apply -f https://k8s.io/examples/debug/counter-pod.yaml

The output is:

pod/counter created

To fetch the logs, use the kubectl logs command, as follows:

kubectl logs counter

The output is similar to:

0: Fri Apr  1 11:42:23 UTC 2022
1: Fri Apr  1 11:42:24 UTC 2022
2: Fri Apr  1 11:42:25 UTC 2022

You can use kubectl logs --previous to retrieve logs from a previous instantiation of a container. If your pod has multiple containers, specify which container's logs you want to access by appending a container name to the command, with a -c flag, like so:

kubectl logs counter -c count

See the kubectl logs documentation for more details.

How nodes handle container logs

Node level logging

A container runtime handles and redirects any output generated to a containerized application's stdout and stderr streams. Different container runtimes implement this in different ways; however, the integration with the kubelet is standardized as the CRI logging format.

By default, if a container restarts, the kubelet keeps one terminated container with its logs. If a pod is evicted from the node, all corresponding containers are also evicted, along with their logs.

The kubelet makes logs available to clients via a special feature of the Kubernetes API. The usual way to access this is by running kubectl logs.

Log rotation

FEATURE STATE: Kubernetes v1.21 [stable]

The kubelet is responsible for rotating container logs and managing the logging directory structure. The kubelet sends this information to the container runtime (using CRI), and the runtime writes the container logs to the given location.

You can configure two kubelet configuration settings, containerLogMaxSize (default 10Mi) and containerLogMaxFiles (default 5), using the kubelet configuration file. These settings let you configure the maximum size for each log file and the maximum number of files allowed for each container respectively.

In order to perform an efficient log rotation in clusters where the volume of the logs generated by the workload is large, kubelet also provides a mechanism to tune how the logs are rotated in terms of how many concurrent log rotations can be performed and the interval at which the logs are monitored and rotated as required. You can configure two kubelet configuration settings, containerLogMaxWorkers and containerLogMonitorInterval using the kubelet configuration file.

When you run kubectl logs as in the basic logging example, the kubelet on the node handles the request and reads directly from the log file. The kubelet returns the content of the log file.

System component logs

There are two types of system components: those that typically run in a container, and those components directly involved in running containers. For example:

  • The kubelet and container runtime do not run in containers. The kubelet runs your containers (grouped together in pods)
  • The Kubernetes scheduler, controller manager, and API server run within pods (usually static Pods). The etcd component runs in the control plane, and most commonly also as a static pod. If your cluster uses kube-proxy, you typically run this as a DaemonSet.

Log locations

The way that the kubelet and container runtime write logs depends on the operating system that the node uses:

On Linux nodes that use systemd, the kubelet and container runtime write to journald by default. You use journalctl to read the systemd journal; for example: journalctl -u kubelet.

If systemd is not present, the kubelet and container runtime write to .log files in the /var/log directory. If you want to have logs written elsewhere, you can indirectly run the kubelet via a helper tool, kube-log-runner, and use that tool to redirect kubelet logs to a directory that you choose.

By default, kubelet directs your container runtime to write logs into directories within /var/log/pods.

For more information on kube-log-runner, read System Logs.

By default, the kubelet writes logs to files within the directory C:\var\logs (notice that this is not C:\var\log).

Although C:\var\log is the Kubernetes default location for these logs, several cluster deployment tools set up Windows nodes to log to C:\var\log\kubelet instead.

If you want to have logs written elsewhere, you can indirectly run the kubelet via a helper tool, kube-log-runner, and use that tool to redirect kubelet logs to a directory that you choose.

However, by default, kubelet directs your container runtime to write logs within the directory C:\var\log\pods.

For more information on kube-log-runner, read System Logs.


For Kubernetes cluster components that run in pods, these write to files inside the /var/log directory, bypassing the default logging mechanism (the components do not write to the systemd journal). You can use Kubernetes' storage mechanisms to map persistent storage into the container that runs the component.

Kubelet allows changing the pod logs directory from default /var/log/pods to a custom path. This adjustment can be made by configuring the podLogsDir parameter in the kubelet's configuration file.

For details about etcd and its logs, view the etcd documentation. Again, you can use Kubernetes' storage mechanisms to map persistent storage into the container that runs the component.

Cluster-level logging architectures

While Kubernetes does not provide a native solution for cluster-level logging, there are several common approaches you can consider. Here are some options:

  • Use a node-level logging agent that runs on every node.
  • Include a dedicated sidecar container for logging in an application pod.
  • Push logs directly to a backend from within an application.

Using a node logging agent

Using a node level logging agent

You can implement cluster-level logging by including a node-level logging agent on each node. The logging agent is a dedicated tool that exposes logs or pushes logs to a backend. Commonly, the logging agent is a container that has access to a directory with log files from all of the application containers on that node.

Because the logging agent must run on every node, it is recommended to run the agent as a DaemonSet.

Node-level logging creates only one agent per node and doesn't require any changes to the applications running on the node.

Containers write to stdout and stderr, but with no agreed format. A node-level agent collects these logs and forwards them for aggregation.

Using a sidecar container with the logging agent

You can use a sidecar container in one of the following ways:

  • The sidecar container streams application logs to its own stdout.
  • The sidecar container runs a logging agent, which is configured to pick up logs from an application container.

Streaming sidecar container

Sidecar container with a streaming container

By having your sidecar containers write to their own stdout and stderr streams, you can take advantage of the kubelet and the logging agent that already run on each node. The sidecar containers read logs from a file, a socket, or journald. Each sidecar container prints a log to its own stdout or stderr stream.

This approach allows you to separate several log streams from different parts of your application, some of which can lack support for writing to stdout or stderr. The logic behind redirecting logs is minimal, so it's not a significant overhead. Additionally, because stdout and stderr are handled by the kubelet, you can use built-in tools like kubectl logs.

For example, a pod runs a single container, and the container writes to two different log files using two different formats. Here's a manifest for the Pod:

apiVersion: v1
kind: Pod
metadata:
  name: counter
spec:
  containers:
  - name: count
    image: busybox:1.28
    args:
    - /bin/sh
    - -c
    - >
      i=0;
      while true;
      do
        echo "$i: $(date)" >> /var/log/1.log;
        echo "$(date) INFO $i" >> /var/log/2.log;
        i=$((i+1));
        sleep 1;
      done      
    volumeMounts:
    - name: varlog
      mountPath: /var/log
  volumes:
  - name: varlog
    emptyDir: {}

It is not recommended to write log entries with different formats to the same log stream, even if you managed to redirect both components to the stdout stream of the container. Instead, you can create two sidecar containers. Each sidecar container could tail a particular log file from a shared volume and then redirect the logs to its own stdout stream.

Here's a manifest for a pod that has two sidecar containers:

apiVersion: v1
kind: Pod
metadata:
  name: counter
spec:
  containers:
  - name: count
    image: busybox:1.28
    args:
    - /bin/sh
    - -c
    - >
      i=0;
      while true;
      do
        echo "$i: $(date)" >> /var/log/1.log;
        echo "$(date) INFO $i" >> /var/log/2.log;
        i=$((i+1));
        sleep 1;
      done      
    volumeMounts:
    - name: varlog
      mountPath: /var/log
  - name: count-log-1
    image: busybox:1.28
    args: [/bin/sh, -c, 'tail -n+1 -F /var/log/1.log']
    volumeMounts:
    - name: varlog
      mountPath: /var/log
  - name: count-log-2
    image: busybox:1.28
    args: [/bin/sh, -c, 'tail -n+1 -F /var/log/2.log']
    volumeMounts:
    - name: varlog
      mountPath: /var/log
  volumes:
  - name: varlog
    emptyDir: {}

Now when you run this pod, you can access each log stream separately by running the following commands:

kubectl logs counter count-log-1

The output is similar to:

0: Fri Apr  1 11:42:26 UTC 2022
1: Fri Apr  1 11:42:27 UTC 2022
2: Fri Apr  1 11:42:28 UTC 2022
...
kubectl logs counter count-log-2

The output is similar to:

Fri Apr  1 11:42:29 UTC 2022 INFO 0
Fri Apr  1 11:42:30 UTC 2022 INFO 0
Fri Apr  1 11:42:31 UTC 2022 INFO 0
...

If you installed a node-level agent in your cluster, that agent picks up those log streams automatically without any further configuration. If you like, you can configure the agent to parse log lines depending on the source container.

Even for Pods that only have low CPU and memory usage (order of a couple of millicores for cpu and order of several megabytes for memory), writing logs to a file and then streaming them to stdout can double how much storage you need on the node. If you have an application that writes to a single file, it's recommended to set /dev/stdout as the destination rather than implement the streaming sidecar container approach.

Sidecar containers can also be used to rotate log files that cannot be rotated by the application itself. An example of this approach is a small container running logrotate periodically. However, it's more straightforward to use stdout and stderr directly, and leave rotation and retention policies to the kubelet.

Sidecar container with a logging agent

Sidecar container with a logging agent

If the node-level logging agent is not flexible enough for your situation, you can create a sidecar container with a separate logging agent that you have configured specifically to run with your application.

Here are two example manifests that you can use to implement a sidecar container with a logging agent. The first manifest contains a ConfigMap to configure fluentd.

apiVersion: v1
kind: ConfigMap
metadata:
  name: fluentd-config
data:
  fluentd.conf: |
    <source>
      type tail
      format none
      path /var/log/1.log
      pos_file /var/log/1.log.pos
      tag count.format1
    </source>

    <source>
      type tail
      format none
      path /var/log/2.log
      pos_file /var/log/2.log.pos
      tag count.format2
    </source>

    <match **>
      type google_cloud
    </match>    

The second manifest describes a pod that has a sidecar container running fluentd. The pod mounts a volume where fluentd can pick up its configuration data.

apiVersion: v1
kind: Pod
metadata:
  name: counter
spec:
  containers:
  - name: count
    image: busybox:1.28
    args:
    - /bin/sh
    - -c
    - >
      i=0;
      while true;
      do
        echo "$i: $(date)" >> /var/log/1.log;
        echo "$(date) INFO $i" >> /var/log/2.log;
        i=$((i+1));
        sleep 1;
      done      
    volumeMounts:
    - name: varlog
      mountPath: /var/log
  - name: count-agent
    image: registry.k8s.io/fluentd-gcp:1.30
    env:
    - name: FLUENTD_ARGS
      value: -c /etc/fluentd-config/fluentd.conf
    volumeMounts:
    - name: varlog
      mountPath: /var/log
    - name: config-volume
      mountPath: /etc/fluentd-config
  volumes:
  - name: varlog
    emptyDir: {}
  - name: config-volume
    configMap:
      name: fluentd-config

Exposing logs directly from the application

Exposing logs directly from the application

Cluster-logging that exposes or pushes logs directly from every application is outside the scope of Kubernetes.

What's next

5 - Metrics For Kubernetes System Components

System component metrics can give a better look into what is happening inside them. Metrics are particularly useful for building dashboards and alerts.

Kubernetes components emit metrics in Prometheus format. This format is structured plain text, designed so that people and machines can both read it.

Metrics in Kubernetes

In most cases metrics are available on /metrics endpoint of the HTTP server. For components that don't expose endpoint by default, it can be enabled using --bind-address flag.

Examples of those components:

In a production environment you may want to configure Prometheus Server or some other metrics scraper to periodically gather these metrics and make them available in some kind of time series database.

Note that kubelet also exposes metrics in /metrics/cadvisor, /metrics/resource and /metrics/probes endpoints. Those metrics do not have the same lifecycle.

If your cluster uses RBAC, reading metrics requires authorization via a user, group or ServiceAccount with a ClusterRole that allows accessing /metrics. For example:

apiVersion: rbac.authorization.k8s.io/v1
kind: ClusterRole
metadata:
  name: prometheus
rules:
  - nonResourceURLs:
      - "/metrics"
    verbs:
      - get

Metric lifecycle

Alpha metric → Stable metric → Deprecated metric → Hidden metric → Deleted metric

Alpha metrics have no stability guarantees. These metrics can be modified or deleted at any time.

Stable metrics are guaranteed to not change. This means:

  • A stable metric without a deprecated signature will not be deleted or renamed
  • A stable metric's type will not be modified

Deprecated metrics are slated for deletion, but are still available for use. These metrics include an annotation about the version in which they became deprecated.

For example:

  • Before deprecation

    # HELP some_counter this counts things
    # TYPE some_counter counter
    some_counter 0
    
  • After deprecation

    # HELP some_counter (Deprecated since 1.15.0) this counts things
    # TYPE some_counter counter
    some_counter 0
    

Hidden metrics are no longer published for scraping, but are still available for use. To use a hidden metric, please refer to the Show hidden metrics section.

Deleted metrics are no longer published and cannot be used.

Show hidden metrics

As described above, admins can enable hidden metrics through a command-line flag on a specific binary. This intends to be used as an escape hatch for admins if they missed the migration of the metrics deprecated in the last release.

The flag show-hidden-metrics-for-version takes a version for which you want to show metrics deprecated in that release. The version is expressed as x.y, where x is the major version, y is the minor version. The patch version is not needed even though a metrics can be deprecated in a patch release, the reason for that is the metrics deprecation policy runs against the minor release.

The flag can only take the previous minor version as it's value. All metrics hidden in previous will be emitted if admins set the previous version to show-hidden-metrics-for-version. The too old version is not allowed because this violates the metrics deprecated policy.

Take metric A as an example, here assumed that A is deprecated in 1.n. According to metrics deprecated policy, we can reach the following conclusion:

  • In release 1.n, the metric is deprecated, and it can be emitted by default.
  • In release 1.n+1, the metric is hidden by default and it can be emitted by command line show-hidden-metrics-for-version=1.n.
  • In release 1.n+2, the metric should be removed from the codebase. No escape hatch anymore.

If you're upgrading from release 1.12 to 1.13, but still depend on a metric A deprecated in 1.12, you should set hidden metrics via command line: --show-hidden-metrics=1.12 and remember to remove this metric dependency before upgrading to 1.14

Component metrics

kube-controller-manager metrics

Controller manager metrics provide important insight into the performance and health of the controller manager. These metrics include common Go language runtime metrics such as go_routine count and controller specific metrics such as etcd request latencies or Cloudprovider (AWS, GCE, OpenStack) API latencies that can be used to gauge the health of a cluster.

Starting from Kubernetes 1.7, detailed Cloudprovider metrics are available for storage operations for GCE, AWS, Vsphere and OpenStack. These metrics can be used to monitor health of persistent volume operations.

For example, for GCE these metrics are called:

cloudprovider_gce_api_request_duration_seconds { request = "instance_list"}
cloudprovider_gce_api_request_duration_seconds { request = "disk_insert"}
cloudprovider_gce_api_request_duration_seconds { request = "disk_delete"}
cloudprovider_gce_api_request_duration_seconds { request = "attach_disk"}
cloudprovider_gce_api_request_duration_seconds { request = "detach_disk"}
cloudprovider_gce_api_request_duration_seconds { request = "list_disk"}

kube-scheduler metrics

FEATURE STATE: Kubernetes v1.21 [beta]

The scheduler exposes optional metrics that reports the requested resources and the desired limits of all running pods. These metrics can be used to build capacity planning dashboards, assess current or historical scheduling limits, quickly identify workloads that cannot schedule due to lack of resources, and compare actual usage to the pod's request.

The kube-scheduler identifies the resource requests and limits configured for each Pod; when either a request or limit is non-zero, the kube-scheduler reports a metrics timeseries. The time series is labelled by:

  • namespace
  • pod name
  • the node where the pod is scheduled or an empty string if not yet scheduled
  • priority
  • the assigned scheduler for that pod
  • the name of the resource (for example, cpu)
  • the unit of the resource if known (for example, cores)

Once a pod reaches completion (has a restartPolicy of Never or OnFailure and is in the Succeeded or Failed pod phase, or has been deleted and all containers have a terminated state) the series is no longer reported since the scheduler is now free to schedule other pods to run. The two metrics are called kube_pod_resource_request and kube_pod_resource_limit.

The metrics are exposed at the HTTP endpoint /metrics/resources and require the same authorization as the /metrics endpoint on the scheduler. You must use the --show-hidden-metrics-for-version=1.20 flag to expose these alpha stability metrics.

Disabling metrics

You can explicitly turn off metrics via command line flag --disabled-metrics. This may be desired if, for example, a metric is causing a performance problem. The input is a list of disabled metrics (i.e. --disabled-metrics=metric1,metric2).

Metric cardinality enforcement

Metrics with unbounded dimensions could cause memory issues in the components they instrument. To limit resource use, you can use the --allow-label-value command line option to dynamically configure an allow-list of label values for a metric.

In alpha stage, the flag can only take in a series of mappings as metric label allow-list. Each mapping is of the format <metric_name>,<label_name>=<allowed_labels> where <allowed_labels> is a comma-separated list of acceptable label names.

The overall format looks like:

--allow-label-value <metric_name>,<label_name>='<allow_value1>, <allow_value2>...', <metric_name2>,<label_name>='<allow_value1>, <allow_value2>...', ...

Here is an example:

--allow-label-value number_count_metric,odd_number='1,3,5', number_count_metric,even_number='2,4,6', date_gauge_metric,weekend='Saturday,Sunday'

In addition to specifying this from the CLI, this can also be done within a configuration file. You can specify the path to that configuration file using the --allow-metric-labels-manifest command line argument to a component. Here's an example of the contents of that configuration file:

allow-list:
- "metric1,label2": "v1,v2,v3"
- "metric2,label1": "v1,v2,v3"

Additionally, the cardinality_enforcement_unexpected_categorizations_total meta-metric records the count of unexpected categorizations during cardinality enforcement, that is, whenever a label value is encountered that is not allowed with respect to the allow-list constraints.

What's next

6 - Metrics for Kubernetes Object States

kube-state-metrics, an add-on agent to generate and expose cluster-level metrics.

The state of Kubernetes objects in the Kubernetes API can be exposed as metrics. An add-on agent called kube-state-metrics can connect to the Kubernetes API server and expose a HTTP endpoint with metrics generated from the state of individual objects in the cluster. It exposes various information about the state of objects like labels and annotations, startup and termination times, status or the phase the object currently is in. For example, containers running in pods create a kube_pod_container_info metric. This includes the name of the container, the name of the pod it is part of, the namespace the pod is running in, the name of the container image, the ID of the image, the image name from the spec of the container, the ID of the running container and the ID of the pod as labels.

An external component that is able and capable to scrape the endpoint of kube-state-metrics (for example via Prometheus) can now be used to enable the following use cases.

Example: using metrics from kube-state-metrics to query the cluster state

Metric series generated by kube-state-metrics are helpful to gather further insights into the cluster, as they can be used for querying.

If you use Prometheus or another tool that uses the same query language, the following PromQL query returns the number of pods that are not ready:

count(kube_pod_status_ready{condition="false"}) by (namespace, pod)

Example: alerting based on from kube-state-metrics

Metrics generated from kube-state-metrics also allow for alerting on issues in the cluster.

If you use Prometheus or a similar tool that uses the same alert rule language, the following alert will fire if there are pods that have been in a Terminating state for more than 5 minutes:

groups:
- name: Pod state
  rules:
  - alert: PodsBlockedInTerminatingState
    expr: count(kube_pod_deletion_timestamp) by (namespace, pod) * count(kube_pod_status_reason{reason="NodeLost"} == 0) by (namespace, pod) > 0
    for: 5m
    labels:
      severity: page
    annotations:
      summary: Pod {{$labels.namespace}}/{{$labels.pod}} blocked in Terminating state.

7 - System Logs

System component logs record events happening in cluster, which can be very useful for debugging. You can configure log verbosity to see more or less detail. Logs can be as coarse-grained as showing errors within a component, or as fine-grained as showing step-by-step traces of events (like HTTP access logs, pod state changes, controller actions, or scheduler decisions).

Klog

klog is the Kubernetes logging library. klog generates log messages for the Kubernetes system components.

Kubernetes is in the process of simplifying logging in its components. The following klog command line flags are deprecated starting with Kubernetes v1.23 and removed in Kubernetes v1.26:

  • --add-dir-header
  • --alsologtostderr
  • --log-backtrace-at
  • --log-dir
  • --log-file
  • --log-file-max-size
  • --logtostderr
  • --one-output
  • --skip-headers
  • --skip-log-headers
  • --stderrthreshold

Output will always be written to stderr, regardless of the output format. Output redirection is expected to be handled by the component which invokes a Kubernetes component. This can be a POSIX shell or a tool like systemd.

In some cases, for example a distroless container or a Windows system service, those options are not available. Then the kube-log-runner binary can be used as wrapper around a Kubernetes component to redirect output. A prebuilt binary is included in several Kubernetes base images under its traditional name as /go-runner and as kube-log-runner in server and node release archives.

This table shows how kube-log-runner invocations correspond to shell redirection:

Usage POSIX shell (such as bash) kube-log-runner <options> <cmd>
Merge stderr and stdout, write to stdout 2>&1 kube-log-runner (default behavior)
Redirect both into log file 1>>/tmp/log 2>&1 kube-log-runner -log-file=/tmp/log
Copy into log file and to stdout 2>&1 | tee -a /tmp/log kube-log-runner -log-file=/tmp/log -also-stdout
Redirect only stdout into log file >/tmp/log kube-log-runner -log-file=/tmp/log -redirect-stderr=false

Klog output

An example of the traditional klog native format:

I1025 00:15:15.525108       1 httplog.go:79] GET /api/v1/namespaces/kube-system/pods/metrics-server-v0.3.1-57c75779f-9p8wg: (1.512ms) 200 [pod_nanny/v0.0.0 (linux/amd64) kubernetes/$Format 10.56.1.19:51756]

The message string may contain line breaks:

I1025 00:15:15.525108       1 example.go:79] This is a message
which has a line break.

Structured Logging

FEATURE STATE: Kubernetes v1.23 [beta]

Structured logging introduces a uniform structure in log messages allowing for programmatic extraction of information. You can store and process structured logs with less effort and cost. The code which generates a log message determines whether it uses the traditional unstructured klog output or structured logging.

The default formatting of structured log messages is as text, with a format that is backward compatible with traditional klog:

<klog header> "<message>" <key1>="<value1>" <key2>="<value2>" ...

Example:

I1025 00:15:15.525108       1 controller_utils.go:116] "Pod status updated" pod="kube-system/kubedns" status="ready"

Strings are quoted. Other values are formatted with %+v, which may cause log messages to continue on the next line depending on the data.

I1025 00:15:15.525108       1 example.go:116] "Example" data="This is text with a line break\nand \"quotation marks\"." someInt=1 someFloat=0.1 someStruct={StringField: First line,
second line.}

Contextual Logging

FEATURE STATE: Kubernetes v1.30 [beta]

Contextual logging builds on top of structured logging. It is primarily about how developers use logging calls: code based on that concept is more flexible and supports additional use cases as described in the Contextual Logging KEP.

If developers use additional functions like WithValues or WithName in their components, then log entries contain additional information that gets passed into functions by their caller.

For Kubernetes 1.30, this is gated behind the ContextualLogging feature gate and is enabled by default. The infrastructure for this was added in 1.24 without modifying components. The component-base/logs/example command demonstrates how to use the new logging calls and how a component behaves that supports contextual logging.

$ cd $GOPATH/src/k8s.io/kubernetes/staging/src/k8s.io/component-base/logs/example/cmd/
$ go run . --help
...
      --feature-gates mapStringBool  A set of key=value pairs that describe feature gates for alpha/experimental features. Options are:
                                     AllAlpha=true|false (ALPHA - default=false)
                                     AllBeta=true|false (BETA - default=false)
                                     ContextualLogging=true|false (BETA - default=true)
$ go run . --feature-gates ContextualLogging=true
...
I0222 15:13:31.645988  197901 example.go:54] "runtime" logger="example.myname" foo="bar" duration="1m0s"
I0222 15:13:31.646007  197901 example.go:55] "another runtime" logger="example" foo="bar" duration="1h0m0s" duration="1m0s"

The logger key and foo="bar" were added by the caller of the function which logs the runtime message and duration="1m0s" value, without having to modify that function.

With contextual logging disable, WithValues and WithName do nothing and log calls go through the global klog logger. Therefore this additional information is not in the log output anymore:

$ go run . --feature-gates ContextualLogging=false
...
I0222 15:14:40.497333  198174 example.go:54] "runtime" duration="1m0s"
I0222 15:14:40.497346  198174 example.go:55] "another runtime" duration="1h0m0s" duration="1m0s"

JSON log format

FEATURE STATE: Kubernetes v1.19 [alpha]

The --logging-format=json flag changes the format of logs from klog native format to JSON format. Example of JSON log format (pretty printed):

{
   "ts": 1580306777.04728,
   "v": 4,
   "msg": "Pod status updated",
   "pod":{
      "name": "nginx-1",
      "namespace": "default"
   },
   "status": "ready"
}

Keys with special meaning:

  • ts - timestamp as Unix time (required, float)
  • v - verbosity (only for info and not for error messages, int)
  • err - error string (optional, string)
  • msg - message (required, string)

List of components currently supporting JSON format:

Log verbosity level

The -v flag controls log verbosity. Increasing the value increases the number of logged events. Decreasing the value decreases the number of logged events. Increasing verbosity settings logs increasingly less severe events. A verbosity setting of 0 logs only critical events.

Log location

There are two types of system components: those that run in a container and those that do not run in a container. For example:

  • The Kubernetes scheduler and kube-proxy run in a container.
  • The kubelet and container runtime do not run in containers.

On machines with systemd, the kubelet and container runtime write to journald. Otherwise, they write to .log files in the /var/log directory. System components inside containers always write to .log files in the /var/log directory, bypassing the default logging mechanism. Similar to the container logs, you should rotate system component logs in the /var/log directory. In Kubernetes clusters created by the kube-up.sh script, log rotation is configured by the logrotate tool. The logrotate tool rotates logs daily, or once the log size is greater than 100MB.

Log query

FEATURE STATE: Kubernetes v1.30 [beta]

To help with debugging issues on nodes, Kubernetes v1.27 introduced a feature that allows viewing logs of services running on the node. To use the feature, ensure that the NodeLogQuery feature gate is enabled for that node, and that the kubelet configuration options enableSystemLogHandler and enableSystemLogQuery are both set to true. On Linux the assumption is that service logs are available via journald. On Windows the assumption is that service logs are available in the application log provider. On both operating systems, logs are also available by reading files within /var/log/.

Provided you are authorized to interact with node objects, you can try out this feature on all your nodes or just a subset. Here is an example to retrieve the kubelet service logs from a node:

# Fetch kubelet logs from a node named node-1.example
kubectl get --raw "/api/v1/nodes/node-1.example/proxy/logs/?query=kubelet"

You can also fetch files, provided that the files are in a directory that the kubelet allows for log fetches. For example, you can fetch a log from /var/log on a Linux node:

kubectl get --raw "/api/v1/nodes/<insert-node-name-here>/proxy/logs/?query=/<insert-log-file-name-here>"

The kubelet uses heuristics to retrieve logs. This helps if you are not aware whether a given system service is writing logs to the operating system's native logger like journald or to a log file in /var/log/. The heuristics first checks the native logger and if that is not available attempts to retrieve the first logs from /var/log/<servicename> or /var/log/<servicename>.log or /var/log/<servicename>/<servicename>.log.

The complete list of options that can be used are:

Option Description
boot boot show messages from a specific system boot
pattern pattern filters log entries by the provided PERL-compatible regular expression
query query specifies services(s) or files from which to return logs (required)
sinceTime an RFC3339 timestamp from which to show logs (inclusive)
untilTime an RFC3339 timestamp until which to show logs (inclusive)
tailLines specify how many lines from the end of the log to retrieve; the default is to fetch the whole log

Example of a more complex query:

# Fetch kubelet logs from a node named node-1.example that have the word "error"
kubectl get --raw "/api/v1/nodes/node-1.example/proxy/logs/?query=kubelet&pattern=error"

What's next

8 - Traces For Kubernetes System Components

FEATURE STATE: Kubernetes v1.27 [beta]

System component traces record the latency of and relationships between operations in the cluster.

Kubernetes components emit traces using the OpenTelemetry Protocol with the gRPC exporter and can be collected and routed to tracing backends using an OpenTelemetry Collector.

Trace Collection

Kubernetes components have built-in gRPC exporters for OTLP to export traces, either with an OpenTelemetry Collector, or without an OpenTelemetry Collector.

For a complete guide to collecting traces and using the collector, see Getting Started with the OpenTelemetry Collector. However, there are a few things to note that are specific to Kubernetes components.

By default, Kubernetes components export traces using the grpc exporter for OTLP on the IANA OpenTelemetry port, 4317. As an example, if the collector is running as a sidecar to a Kubernetes component, the following receiver configuration will collect spans and log them to standard output:

receivers:
  otlp:
    protocols:
      grpc:
exporters:
  # Replace this exporter with the exporter for your backend
  logging:
    logLevel: debug
service:
  pipelines:
    traces:
      receivers: [otlp]
      exporters: [logging]

To directly emit traces to a backend without utilizing a collector, specify the endpoint field in the Kubernetes tracing configuration file with the desired trace backend address. This method negates the need for a collector and simplifies the overall structure.

For trace backend header configuration, including authentication details, environment variables can be used with OTEL_EXPORTER_OTLP_HEADERS, see OTLP Exporter Configuration.

Additionally, for trace resource attribute configuration such as Kubernetes cluster name, namespace, Pod name, etc., environment variables can also be used with OTEL_RESOURCE_ATTRIBUTES, see OTLP Kubernetes Resource.

Component traces

kube-apiserver traces

The kube-apiserver generates spans for incoming HTTP requests, and for outgoing requests to webhooks, etcd, and re-entrant requests. It propagates the W3C Trace Context with outgoing requests but does not make use of the trace context attached to incoming requests, as the kube-apiserver is often a public endpoint.

Enabling tracing in the kube-apiserver

To enable tracing, provide the kube-apiserver with a tracing configuration file with --tracing-config-file=<path-to-config>. This is an example config that records spans for 1 in 10000 requests, and uses the default OpenTelemetry endpoint:

apiVersion: apiserver.config.k8s.io/v1beta1
kind: TracingConfiguration
# default value
#endpoint: localhost:4317
samplingRatePerMillion: 100

For more information about the TracingConfiguration struct, see API server config API (v1beta1).

kubelet traces

FEATURE STATE: Kubernetes v1.27 [beta]

The kubelet CRI interface and authenticated http servers are instrumented to generate trace spans. As with the apiserver, the endpoint and sampling rate are configurable. Trace context propagation is also configured. A parent span's sampling decision is always respected. A provided tracing configuration sampling rate will apply to spans without a parent. Enabled without a configured endpoint, the default OpenTelemetry Collector receiver address of "localhost:4317" is set.

Enabling tracing in the kubelet

To enable tracing, apply the tracing configuration. This is an example snippet of a kubelet config that records spans for 1 in 10000 requests, and uses the default OpenTelemetry endpoint:

apiVersion: kubelet.config.k8s.io/v1beta1
kind: KubeletConfiguration
featureGates:
  KubeletTracing: true
tracing:
  # default value
  #endpoint: localhost:4317
  samplingRatePerMillion: 100

If the samplingRatePerMillion is set to one million (1000000), then every span will be sent to the exporter.

The kubelet in Kubernetes v1.30 collects spans from the garbage collection, pod synchronization routine as well as every gRPC method. The kubelet propagates trace context with gRPC requests so that container runtimes with trace instrumentation, such as CRI-O and containerd, can associate their exported spans with the trace context from the kubelet. The resulting traces will have parent-child links between kubelet and container runtime spans, providing helpful context when debugging node issues.

Please note that exporting spans always comes with a small performance overhead on the networking and CPU side, depending on the overall configuration of the system. If there is any issue like that in a cluster which is running with tracing enabled, then mitigate the problem by either reducing the samplingRatePerMillion or disabling tracing completely by removing the configuration.

Stability

Tracing instrumentation is still under active development, and may change in a variety of ways. This includes span names, attached attributes, instrumented endpoints, etc. Until this feature graduates to stable, there are no guarantees of backwards compatibility for tracing instrumentation.

What's next

9 - Proxies in Kubernetes

This page explains proxies used with Kubernetes.

Proxies

There are several different proxies you may encounter when using Kubernetes:

  1. The kubectl proxy:

    • runs on a user's desktop or in a pod
    • proxies from a localhost address to the Kubernetes apiserver
    • client to proxy uses HTTP
    • proxy to apiserver uses HTTPS
    • locates apiserver
    • adds authentication headers
  2. The apiserver proxy:

    • is a bastion built into the apiserver
    • connects a user outside of the cluster to cluster IPs which otherwise might not be reachable
    • runs in the apiserver processes
    • client to proxy uses HTTPS (or http if apiserver so configured)
    • proxy to target may use HTTP or HTTPS as chosen by proxy using available information
    • can be used to reach a Node, Pod, or Service
    • does load balancing when used to reach a Service
  3. The kube proxy:

    • runs on each node
    • proxies UDP, TCP and SCTP
    • does not understand HTTP
    • provides load balancing
    • is only used to reach services
  4. A Proxy/Load-balancer in front of apiserver(s):

    • existence and implementation varies from cluster to cluster (e.g. nginx)
    • sits between all clients and one or more apiservers
    • acts as load balancer if there are several apiservers.
  5. Cloud Load Balancers on external services:

    • are provided by some cloud providers (e.g. AWS ELB, Google Cloud Load Balancer)
    • are created automatically when the Kubernetes service has type LoadBalancer
    • usually supports UDP/TCP only
    • SCTP support is up to the load balancer implementation of the cloud provider
    • implementation varies by cloud provider.

Kubernetes users will typically not need to worry about anything other than the first two types. The cluster admin will typically ensure that the latter types are set up correctly.

Requesting redirects

Proxies have replaced redirect capabilities. Redirects have been deprecated.

10 - API Priority and Fairness

FEATURE STATE: Kubernetes v1.29 [stable]

Controlling the behavior of the Kubernetes API server in an overload situation is a key task for cluster administrators. The kube-apiserver has some controls available (i.e. the --max-requests-inflight and --max-mutating-requests-inflight command-line flags) to limit the amount of outstanding work that will be accepted, preventing a flood of inbound requests from overloading and potentially crashing the API server, but these flags are not enough to ensure that the most important requests get through in a period of high traffic.

The API Priority and Fairness feature (APF) is an alternative that improves upon aforementioned max-inflight limitations. APF classifies and isolates requests in a more fine-grained way. It also introduces a limited amount of queuing, so that no requests are rejected in cases of very brief bursts. Requests are dispatched from queues using a fair queuing technique so that, for example, a poorly-behaved controller need not starve others (even at the same priority level).

This feature is designed to work well with standard controllers, which use informers and react to failures of API requests with exponential back-off, and other clients that also work this way.

Enabling/Disabling API Priority and Fairness

The API Priority and Fairness feature is controlled by a command-line flag and is enabled by default. See Options for a general explanation of the available kube-apiserver command-line options and how to enable and disable them. The name of the command-line option for APF is "--enable-priority-and-fairness". This feature also involves an API Group with: (a) a stable v1 version, introduced in 1.29, and enabled by default (b) a v1beta3 version, enabled by default, and deprecated in v1.29. You can disable the API group beta version v1beta3 by adding the following command-line flags to your kube-apiserver invocation:

kube-apiserver \
--runtime-config=flowcontrol.apiserver.k8s.io/v1beta3=false \
 # …and other flags as usual

The command-line flag --enable-priority-and-fairness=false will disable the API Priority and Fairness feature.

Recursive server scenarios

API Priority and Fairness must be used carefully in recursive server scenarios. These are scenarios in which some server A, while serving a request, issues a subsidiary request to some server B. Perhaps server B might even make a further subsidiary call back to server A. In situations where Priority and Fairness control is applied to both the original request and some subsidiary ones(s), no matter how deep in the recursion, there is a danger of priority inversions and/or deadlocks.

One example of recursion is when the kube-apiserver issues an admission webhook call to server B, and while serving that call, server B makes a further subsidiary request back to the kube-apiserver. Another example of recursion is when an APIService object directs the kube-apiserver to delegate requests about a certain API group to a custom external server B (this is one of the things called "aggregation").

When the original request is known to belong to a certain priority level, and the subsidiary controlled requests are classified to higher priority levels, this is one possible solution. When the original requests can belong to any priority level, the subsidiary controlled requests have to be exempt from Priority and Fairness limitation. One way to do that is with the objects that configure classification and handling, discussed below. Another way is to disable Priority and Fairness on server B entirely, using the techniques discussed above. A third way, which is the simplest to use when server B is not kube-apisever, is to build server B with Priority and Fairness disabled in the code.

Concepts

There are several distinct features involved in the API Priority and Fairness feature. Incoming requests are classified by attributes of the request using FlowSchemas, and assigned to priority levels. Priority levels add a degree of isolation by maintaining separate concurrency limits, so that requests assigned to different priority levels cannot starve each other. Within a priority level, a fair-queuing algorithm prevents requests from different flows from starving each other, and allows for requests to be queued to prevent bursty traffic from causing failed requests when the average load is acceptably low.

Priority Levels

Without APF enabled, overall concurrency in the API server is limited by the kube-apiserver flags --max-requests-inflight and --max-mutating-requests-inflight. With APF enabled, the concurrency limits defined by these flags are summed and then the sum is divided up among a configurable set of priority levels. Each incoming request is assigned to a single priority level, and each priority level will only dispatch as many concurrent requests as its particular limit allows.

The default configuration, for example, includes separate priority levels for leader-election requests, requests from built-in controllers, and requests from Pods. This means that an ill-behaved Pod that floods the API server with requests cannot prevent leader election or actions by the built-in controllers from succeeding.

The concurrency limits of the priority levels are periodically adjusted, allowing under-utilized priority levels to temporarily lend concurrency to heavily-utilized levels. These limits are based on nominal limits and bounds on how much concurrency a priority level may lend and how much it may borrow, all derived from the configuration objects mentioned below.

Seats Occupied by a Request

The above description of concurrency management is the baseline story. Requests have different durations but are counted equally at any given moment when comparing against a priority level's concurrency limit. In the baseline story, each request occupies one unit of concurrency. The word "seat" is used to mean one unit of concurrency, inspired by the way each passenger on a train or aircraft takes up one of the fixed supply of seats.

But some requests take up more than one seat. Some of these are list requests that the server estimates will return a large number of objects. These have been found to put an exceptionally heavy burden on the server. For this reason, the server estimates the number of objects that will be returned and considers the request to take a number of seats that is proportional to that estimated number.

Execution time tweaks for watch requests

API Priority and Fairness manages watch requests, but this involves a couple more excursions from the baseline behavior. The first concerns how long a watch request is considered to occupy its seat. Depending on request parameters, the response to a watch request may or may not begin with create notifications for all the relevant pre-existing objects. API Priority and Fairness considers a watch request to be done with its seat once that initial burst of notifications, if any, is over.

The normal notifications are sent in a concurrent burst to all relevant watch response streams whenever the server is notified of an object create/update/delete. To account for this work, API Priority and Fairness considers every write request to spend some additional time occupying seats after the actual writing is done. The server estimates the number of notifications to be sent and adjusts the write request's number of seats and seat occupancy time to include this extra work.

Queuing

Even within a priority level there may be a large number of distinct sources of traffic. In an overload situation, it is valuable to prevent one stream of requests from starving others (in particular, in the relatively common case of a single buggy client flooding the kube-apiserver with requests, that buggy client would ideally not have much measurable impact on other clients at all). This is handled by use of a fair-queuing algorithm to process requests that are assigned the same priority level. Each request is assigned to a flow, identified by the name of the matching FlowSchema plus a flow distinguisher — which is either the requesting user, the target resource's namespace, or nothing — and the system attempts to give approximately equal weight to requests in different flows of the same priority level. To enable distinct handling of distinct instances, controllers that have many instances should authenticate with distinct usernames

After classifying a request into a flow, the API Priority and Fairness feature then may assign the request to a queue. This assignment uses a technique known as shuffle sharding, which makes relatively efficient use of queues to insulate low-intensity flows from high-intensity flows.

The details of the queuing algorithm are tunable for each priority level, and allow administrators to trade off memory use, fairness (the property that independent flows will all make progress when total traffic exceeds capacity), tolerance for bursty traffic, and the added latency induced by queuing.

Exempt requests

Some requests are considered sufficiently important that they are not subject to any of the limitations imposed by this feature. These exemptions prevent an improperly-configured flow control configuration from totally disabling an API server.

Resources

The flow control API involves two kinds of resources. PriorityLevelConfigurations define the available priority levels, the share of the available concurrency budget that each can handle, and allow for fine-tuning queuing behavior. FlowSchemas are used to classify individual inbound requests, matching each to a single PriorityLevelConfiguration.

PriorityLevelConfiguration

A PriorityLevelConfiguration represents a single priority level. Each PriorityLevelConfiguration has an independent limit on the number of outstanding requests, and limitations on the number of queued requests.

The nominal concurrency limit for a PriorityLevelConfiguration is not specified in an absolute number of seats, but rather in "nominal concurrency shares." The total concurrency limit for the API Server is distributed among the existing PriorityLevelConfigurations in proportion to these shares, to give each level its nominal limit in terms of seats. This allows a cluster administrator to scale up or down the total amount of traffic to a server by restarting kube-apiserver with a different value for --max-requests-inflight (or --max-mutating-requests-inflight), and all PriorityLevelConfigurations will see their maximum allowed concurrency go up (or down) by the same fraction.

The bounds on how much concurrency a priority level may lend and how much it may borrow are expressed in the PriorityLevelConfiguration as percentages of the level's nominal limit. These are resolved to absolute numbers of seats by multiplying with the nominal limit / 100.0 and rounding. The dynamically adjusted concurrency limit of a priority level is constrained to lie between (a) a lower bound of its nominal limit minus its lendable seats and (b) an upper bound of its nominal limit plus the seats it may borrow. At each adjustment the dynamic limits are derived by each priority level reclaiming any lent seats for which demand recently appeared and then jointly fairly responding to the recent seat demand on the priority levels, within the bounds just described.

When the volume of inbound requests assigned to a single PriorityLevelConfiguration is more than its permitted concurrency level, the type field of its specification determines what will happen to extra requests. A type of Reject means that excess traffic will immediately be rejected with an HTTP 429 (Too Many Requests) error. A type of Queue means that requests above the threshold will be queued, with the shuffle sharding and fair queuing techniques used to balance progress between request flows.

The queuing configuration allows tuning the fair queuing algorithm for a priority level. Details of the algorithm can be read in the enhancement proposal, but in short:

  • Increasing queues reduces the rate of collisions between different flows, at the cost of increased memory usage. A value of 1 here effectively disables the fair-queuing logic, but still allows requests to be queued.

  • Increasing queueLengthLimit allows larger bursts of traffic to be sustained without dropping any requests, at the cost of increased latency and memory usage.

  • Changing handSize allows you to adjust the probability of collisions between different flows and the overall concurrency available to a single flow in an overload situation.

Following is a table showing an interesting collection of shuffle sharding configurations, showing for each the probability that a given mouse (low-intensity flow) is squished by the elephants (high-intensity flows) for an illustrative collection of numbers of elephants. See https://play.golang.org/p/Gi0PLgVHiUg , which computes this table.

Example Shuffle Sharding Configurations
HandSize Queues 1 elephant 4 elephants 16 elephants
12 32 4.428838398950118e-09 0.11431348830099144 0.9935089607656024
10 32 1.550093439632541e-08 0.0626479840223545 0.9753101519027554
10 64 6.601827268370426e-12 0.00045571320990370776 0.49999929150089345
9 64 3.6310049976037345e-11 0.00045501212304112273 0.4282314876454858
8 64 2.25929199850899e-10 0.0004886697053040446 0.35935114681123076
8 128 6.994461389026097e-13 3.4055790161620863e-06 0.02746173137155063
7 128 1.0579122850901972e-11 6.960839379258192e-06 0.02406157386340147
7 256 7.597695465552631e-14 6.728547142019406e-08 0.0006709661542533682
6 256 2.7134626662687968e-12 2.9516464018476436e-07 0.0008895654642000348
6 512 4.116062922897309e-14 4.982983350480894e-09 2.26025764343413e-05
6 1024 6.337324016514285e-16 8.09060164312957e-11 4.517408062903668e-07

FlowSchema

A FlowSchema matches some inbound requests and assigns them to a priority level. Every inbound request is tested against FlowSchemas, starting with those with the numerically lowest matchingPrecedence and working upward. The first match wins.

A FlowSchema matches a given request if at least one of its rules matches. A rule matches if at least one of its subjects and at least one of its resourceRules or nonResourceRules (depending on whether the incoming request is for a resource or non-resource URL) match the request.

For the name field in subjects, and the verbs, apiGroups, resources, namespaces, and nonResourceURLs fields of resource and non-resource rules, the wildcard * may be specified to match all values for the given field, effectively removing it from consideration.

A FlowSchema's distinguisherMethod.type determines how requests matching that schema will be separated into flows. It may be ByUser, in which one requesting user will not be able to starve other users of capacity; ByNamespace, in which requests for resources in one namespace will not be able to starve requests for resources in other namespaces of capacity; or blank (or distinguisherMethod may be omitted entirely), in which all requests matched by this FlowSchema will be considered part of a single flow. The correct choice for a given FlowSchema depends on the resource and your particular environment.

Defaults

Each kube-apiserver maintains two sorts of APF configuration objects: mandatory and suggested.

Mandatory Configuration Objects

The four mandatory configuration objects reflect fixed built-in guardrail behavior. This is behavior that the servers have before those objects exist, and when those objects exist their specs reflect this behavior. The four mandatory objects are as follows.

  • The mandatory exempt priority level is used for requests that are not subject to flow control at all: they will always be dispatched immediately. The mandatory exempt FlowSchema classifies all requests from the system:masters group into this priority level. You may define other FlowSchemas that direct other requests to this priority level, if appropriate.

  • The mandatory catch-all priority level is used in combination with the mandatory catch-all FlowSchema to make sure that every request gets some kind of classification. Typically you should not rely on this catch-all configuration, and should create your own catch-all FlowSchema and PriorityLevelConfiguration (or use the suggested global-default priority level that is installed by default) as appropriate. Because it is not expected to be used normally, the mandatory catch-all priority level has a very small concurrency share and does not queue requests.

Suggested Configuration Objects

The suggested FlowSchemas and PriorityLevelConfigurations constitute a reasonable default configuration. You can modify these and/or create additional configuration objects if you want. If your cluster is likely to experience heavy load then you should consider what configuration will work best.

The suggested configuration groups requests into six priority levels:

  • The node-high priority level is for health updates from nodes.

  • The system priority level is for non-health requests from the system:nodes group, i.e. Kubelets, which must be able to contact the API server in order for workloads to be able to schedule on them.

  • The leader-election priority level is for leader election requests from built-in controllers (in particular, requests for endpoints, configmaps, or leases coming from the system:kube-controller-manager or system:kube-scheduler users and service accounts in the kube-system namespace). These are important to isolate from other traffic because failures in leader election cause their controllers to fail and restart, which in turn causes more expensive traffic as the new controllers sync their informers.

  • The workload-high priority level is for other requests from built-in controllers.

  • The workload-low priority level is for requests from any other service account, which will typically include all requests from controllers running in Pods.

  • The global-default priority level handles all other traffic, e.g. interactive kubectl commands run by nonprivileged users.

The suggested FlowSchemas serve to steer requests into the above priority levels, and are not enumerated here.

Maintenance of the Mandatory and Suggested Configuration Objects

Each kube-apiserver independently maintains the mandatory and suggested configuration objects, using initial and periodic behavior. Thus, in a situation with a mixture of servers of different versions there may be thrashing as long as different servers have different opinions of the proper content of these objects.

Each kube-apiserver makes an initial maintenance pass over the mandatory and suggested configuration objects, and after that does periodic maintenance (once per minute) of those objects.

For the mandatory configuration objects, maintenance consists of ensuring that the object exists and, if it does, has the proper spec. The server refuses to allow a creation or update with a spec that is inconsistent with the server's guardrail behavior.

Maintenance of suggested configuration objects is designed to allow their specs to be overridden. Deletion, on the other hand, is not respected: maintenance will restore the object. If you do not want a suggested configuration object then you need to keep it around but set its spec to have minimal consequences. Maintenance of suggested objects is also designed to support automatic migration when a new version of the kube-apiserver is rolled out, albeit potentially with thrashing while there is a mixed population of servers.

Maintenance of a suggested configuration object consists of creating it --- with the server's suggested spec --- if the object does not exist. OTOH, if the object already exists, maintenance behavior depends on whether the kube-apiservers or the users control the object. In the former case, the server ensures that the object's spec is what the server suggests; in the latter case, the spec is left alone.

The question of who controls the object is answered by first looking for an annotation with key apf.kubernetes.io/autoupdate-spec. If there is such an annotation and its value is true then the kube-apiservers control the object. If there is such an annotation and its value is false then the users control the object. If neither of those conditions holds then the metadata.generation of the object is consulted. If that is 1 then the kube-apiservers control the object. Otherwise the users control the object. These rules were introduced in release 1.22 and their consideration of metadata.generation is for the sake of migration from the simpler earlier behavior. Users who wish to control a suggested configuration object should set its apf.kubernetes.io/autoupdate-spec annotation to false.

Maintenance of a mandatory or suggested configuration object also includes ensuring that it has an apf.kubernetes.io/autoupdate-spec annotation that accurately reflects whether the kube-apiservers control the object.

Maintenance also includes deleting objects that are neither mandatory nor suggested but are annotated apf.kubernetes.io/autoupdate-spec=true.

Health check concurrency exemption

The suggested configuration gives no special treatment to the health check requests on kube-apiservers from their local kubelets --- which tend to use the secured port but supply no credentials. With the suggested config, these requests get assigned to the global-default FlowSchema and the corresponding global-default priority level, where other traffic can crowd them out.

If you add the following additional FlowSchema, this exempts those requests from rate limiting.

apiVersion: flowcontrol.apiserver.k8s.io/v1
kind: FlowSchema
metadata:
  name: health-for-strangers
spec:
  matchingPrecedence: 1000
  priorityLevelConfiguration:
    name: exempt
  rules:
    - nonResourceRules:
      - nonResourceURLs:
          - "/healthz"
          - "/livez"
          - "/readyz"
        verbs:
          - "*"
      subjects:
        - kind: Group
          group:
            name: "system:unauthenticated"

Observability

Metrics

When you enable the API Priority and Fairness feature, the kube-apiserver exports additional metrics. Monitoring these can help you determine whether your configuration is inappropriately throttling important traffic, or find poorly-behaved workloads that may be harming system health.

Maturity level BETA

  • apiserver_flowcontrol_rejected_requests_total is a counter vector (cumulative since server start) of requests that were rejected, broken down by the labels flow_schema (indicating the one that matched the request), priority_level (indicating the one to which the request was assigned), and reason. The reason label will be one of the following values:

    • queue-full, indicating that too many requests were already queued.
    • concurrency-limit, indicating that the PriorityLevelConfiguration is configured to reject rather than queue excess requests.
    • time-out, indicating that the request was still in the queue when its queuing time limit expired.
    • cancelled, indicating that the request is not purge locked and has been ejected from the queue.
  • apiserver_flowcontrol_dispatched_requests_total is a counter vector (cumulative since server start) of requests that began executing, broken down by flow_schema and priority_level.

  • apiserver_flowcontrol_current_inqueue_requests is a gauge vector holding the instantaneous number of queued (not executing) requests, broken down by priority_level and flow_schema.

  • apiserver_flowcontrol_current_executing_requests is a gauge vector holding the instantaneous number of executing (not waiting in a queue) requests, broken down by priority_level and flow_schema.

  • apiserver_flowcontrol_current_executing_seats is a gauge vector holding the instantaneous number of occupied seats, broken down by priority_level and flow_schema.

  • apiserver_flowcontrol_request_wait_duration_seconds is a histogram vector of how long requests spent queued, broken down by the labels flow_schema, priority_level, and execute. The execute label indicates whether the request has started executing.

  • apiserver_flowcontrol_nominal_limit_seats is a gauge vector holding each priority level's nominal concurrency limit, computed from the API server's total concurrency limit and the priority level's configured nominal concurrency shares.

Maturity level ALPHA

  • apiserver_current_inqueue_requests is a gauge vector of recent high water marks of the number of queued requests, grouped by a label named request_kind whose value is mutating or readOnly. These high water marks describe the largest number seen in the one second window most recently completed. These complement the older apiserver_current_inflight_requests gauge vector that holds the last window's high water mark of number of requests actively being served.

  • apiserver_current_inqueue_seats is a gauge vector of the sum over queued requests of the largest number of seats each will occupy, grouped by labels named flow_schema and priority_level.

  • apiserver_flowcontrol_read_vs_write_current_requests is a histogram vector of observations, made at the end of every nanosecond, of the number of requests broken down by the labels phase (which takes on the values waiting and executing) and request_kind (which takes on the values mutating and readOnly). Each observed value is a ratio, between 0 and 1, of the number of requests divided by the corresponding limit on the number of requests (queue volume limit for waiting and concurrency limit for executing).

  • apiserver_flowcontrol_request_concurrency_in_use is a gauge vector holding the instantaneous number of occupied seats, broken down by priority_level and flow_schema.

  • apiserver_flowcontrol_priority_level_request_utilization is a histogram vector of observations, made at the end of each nanosecond, of the number of requests broken down by the labels phase (which takes on the values waiting and executing) and priority_level. Each observed value is a ratio, between 0 and 1, of a number of requests divided by the corresponding limit on the number of requests (queue volume limit for waiting and concurrency limit for executing).

  • apiserver_flowcontrol_priority_level_seat_utilization is a histogram vector of observations, made at the end of each nanosecond, of the utilization of a priority level's concurrency limit, broken down by priority_level. This utilization is the fraction (number of seats occupied) / (concurrency limit). This metric considers all stages of execution (both normal and the extra delay at the end of a write to cover for the corresponding notification work) of all requests except WATCHes; for those it considers only the initial stage that delivers notifications of pre-existing objects. Each histogram in the vector is also labeled with phase: executing (there is no seat limit for the waiting phase).

  • apiserver_flowcontrol_request_queue_length_after_enqueue is a histogram vector of queue lengths for the queues, broken down by priority_level and flow_schema, as sampled by the enqueued requests. Each request that gets queued contributes one sample to its histogram, reporting the length of the queue immediately after the request was added. Note that this produces different statistics than an unbiased survey would.

  • apiserver_flowcontrol_request_concurrency_limit is the same as apiserver_flowcontrol_nominal_limit_seats. Before the introduction of concurrency borrowing between priority levels, this was always equal to apiserver_flowcontrol_current_limit_seats (which did not exist as a distinct metric).

  • apiserver_flowcontrol_lower_limit_seats is a gauge vector holding the lower bound on each priority level's dynamic concurrency limit.

  • apiserver_flowcontrol_upper_limit_seats is a gauge vector holding the upper bound on each priority level's dynamic concurrency limit.

  • apiserver_flowcontrol_demand_seats is a histogram vector counting observations, at the end of every nanosecond, of each priority level's ratio of (seat demand) / (nominal concurrency limit). A priority level's seat demand is the sum, over both queued requests and those in the initial phase of execution, of the maximum of the number of seats occupied in the request's initial and final execution phases.

  • apiserver_flowcontrol_demand_seats_high_watermark is a gauge vector holding, for each priority level, the maximum seat demand seen during the last concurrency borrowing adjustment period.

  • apiserver_flowcontrol_demand_seats_average is a gauge vector holding, for each priority level, the time-weighted average seat demand seen during the last concurrency borrowing adjustment period.

  • apiserver_flowcontrol_demand_seats_stdev is a gauge vector holding, for each priority level, the time-weighted population standard deviation of seat demand seen during the last concurrency borrowing adjustment period.

  • apiserver_flowcontrol_demand_seats_smoothed is a gauge vector holding, for each priority level, the smoothed enveloped seat demand determined at the last concurrency adjustment.

  • apiserver_flowcontrol_target_seats is a gauge vector holding, for each priority level, the concurrency target going into the borrowing allocation problem.

  • apiserver_flowcontrol_seat_fair_frac is a gauge holding the fair allocation fraction determined in the last borrowing adjustment.

  • apiserver_flowcontrol_current_limit_seats is a gauge vector holding, for each priority level, the dynamic concurrency limit derived in the last adjustment.

  • apiserver_flowcontrol_request_execution_seconds is a histogram vector of how long requests took to actually execute, broken down by flow_schema and priority_level.

  • apiserver_flowcontrol_watch_count_samples is a histogram vector of the number of active WATCH requests relevant to a given write, broken down by flow_schema and priority_level.

  • apiserver_flowcontrol_work_estimated_seats is a histogram vector of the number of estimated seats (maximum of initial and final stage of execution) associated with requests, broken down by flow_schema and priority_level.

  • apiserver_flowcontrol_request_dispatch_no_accommodation_total is a counter vector of the number of events that in principle could have led to a request being dispatched but did not, due to lack of available concurrency, broken down by flow_schema and priority_level.

  • apiserver_flowcontrol_epoch_advance_total is a counter vector of the number of attempts to jump a priority level's progress meter backward to avoid numeric overflow, grouped by priority_level and success.

Good practices for using API Priority and Fairness

When a given priority level exceeds its permitted concurrency, requests can experience increased latency or be dropped with an HTTP 429 (Too Many Requests) error. To prevent these side effects of APF, you can modify your workload or tweak your APF settings to ensure there are sufficient seats available to serve your requests.

To detect whether requests are being rejected due to APF, check the following metrics:

  • apiserver_flowcontrol_rejected_requests_total: the total number of requests rejected per FlowSchema and PriorityLevelConfiguration.
  • apiserver_flowcontrol_current_inqueue_requests: the current number of requests queued per FlowSchema and PriorityLevelConfiguration.
  • apiserver_flowcontrol_request_wait_duration_seconds: the latency added to requests waiting in queues.
  • apiserver_flowcontrol_priority_level_seat_utilization: the seat utilization per PriorityLevelConfiguration.

Workload modifications

To prevent requests from queuing and adding latency or being dropped due to APF, you can optimize your requests by:

  • Reducing the rate at which requests are executed. A fewer number of requests over a fixed period will result in a fewer number of seats being needed at a given time.
  • Avoid issuing a large number of expensive requests concurrently. Requests can be optimized to use fewer seats or have lower latency so that these requests hold those seats for a shorter duration. List requests can occupy more than 1 seat depending on the number of objects fetched during the request. Restricting the number of objects retrieved in a list request, for example by using pagination, will use less total seats over a shorter period. Furthermore, replacing list requests with watch requests will require lower total concurrency shares as watch requests only occupy 1 seat during its initial burst of notifications. If using streaming lists in versions 1.27 and later, watch requests will occupy the same number of seats as a list request for its initial burst of notifications because the entire state of the collection has to be streamed. Note that in both cases, a watch request will not hold any seats after this initial phase.

Keep in mind that queuing or rejected requests from APF could be induced by either an increase in the number of requests or an increase in latency for existing requests. For example, if requests that normally take 1s to execute start taking 60s, it is possible that APF will start rejecting requests because requests are occupying seats for a longer duration than normal due to this increase in latency. If APF starts rejecting requests across multiple priority levels without a significant change in workload, it is possible there is an underlying issue with control plane performance rather than the workload or APF settings.

Priority and fairness settings

You can also modify the default FlowSchema and PriorityLevelConfiguration objects or create new objects of these types to better accommodate your workload.

APF settings can be modified to:

  • Give more seats to high priority requests.
  • Isolate non-essential or expensive requests that would starve a concurrency level if it was shared with other flows.

Give more seats to high priority requests

  1. If possible, the number of seats available across all priority levels for a particular kube-apiserver can be increased by increasing the values for the max-requests-inflight and max-mutating-requests-inflight flags. Alternatively, horizontally scaling the number of kube-apiserver instances will increase the total concurrency per priority level across the cluster assuming there is sufficient load balancing of requests.
  2. You can create a new FlowSchema which references a PriorityLevelConfiguration with a larger concurrency level. This new PriorityLevelConfiguration could be an existing level or a new level with its own set of nominal concurrency shares. For example, a new FlowSchema could be introduced to change the PriorityLevelConfiguration for your requests from global-default to workload-low to increase the number of seats available to your user. Creating a new PriorityLevelConfiguration will reduce the number of seats designated for existing levels. Recall that editing a default FlowSchema or PriorityLevelConfiguration will require setting the apf.kubernetes.io/autoupdate-spec annotation to false.
  3. You can also increase the NominalConcurrencyShares for the PriorityLevelConfiguration which is serving your high priority requests. Alternatively, for versions 1.26 and later, you can increase the LendablePercent for competing priority levels so that the given priority level has a higher pool of seats it can borrow.

Isolate non-essential requests from starving other flows

For request isolation, you can create a FlowSchema whose subject matches the user making these requests or create a FlowSchema that matches what the request is (corresponding to the resourceRules). Next, you can map this FlowSchema to a PriorityLevelConfiguration with a low share of seats.

For example, suppose list event requests from Pods running in the default namespace are using 10 seats each and execute for 1 minute. To prevent these expensive requests from impacting requests from other Pods using the existing service-accounts FlowSchema, you can apply the following FlowSchema to isolate these list calls from other requests.

Example FlowSchema object to isolate list event requests:

apiVersion: flowcontrol.apiserver.k8s.io/v1
kind: FlowSchema
metadata:
  name: list-events-default-service-account
spec:
  distinguisherMethod:
    type: ByUser
  matchingPrecedence: 8000
  priorityLevelConfiguration:
    name: catch-all
  rules:
    - resourceRules:
      - apiGroups:
          - '*'
        namespaces:
          - default
        resources:
          - events
        verbs:
          - list
      subjects:
        - kind: ServiceAccount
          serviceAccount:
            name: default
            namespace: default
  • This FlowSchema captures all list event calls made by the default service account in the default namespace. The matching precedence 8000 is lower than the value of 9000 used by the existing service-accounts FlowSchema so these list event calls will match list-events-default-service-account rather than service-accounts.
  • The catch-all PriorityLevelConfiguration is used to isolate these requests. The catch-all priority level has a very small concurrency share and does not queue requests.

What's next

11 - Cluster Autoscaling

Automatically manage the nodes in your cluster to adapt to demand.

Kubernetes requires nodes in your cluster to run pods. This means providing capacity for the workload Pods and for Kubernetes itself.

You can adjust the amount of resources available in your cluster automatically: node autoscaling. You can either change the number of nodes, or change the capacity that nodes provide. The first approach is referred to as horizontal scaling, while the second is referred to as vertical scaling.

Kubernetes can even provide multidimensional automatic scaling for nodes.

Manual node management

You can manually manage node-level capacity, where you configure a fixed amount of nodes; you can use this approach even if the provisioning (the process to set up, manage, and decommission) for these nodes is automated.

This page is about taking the next step, and automating management of the amount of node capacity (CPU, memory, and other node resources) available in your cluster.

Automatic horizontal scaling

Cluster Autoscaler

You can use the Cluster Autoscaler to manage the scale of your nodes automatically. The cluster autoscaler can integrate with a cloud provider, or with Kubernetes' cluster API, to achieve the actual node management that's needed.

The cluster autoscaler adds nodes when there are unschedulable Pods, and removes nodes when those nodes are empty.

Cloud provider integrations

The README for the cluster autoscaler lists some of the cloud provider integrations that are available.

Cost-aware multidimensional scaling

Karpenter

Karpenter supports direct node management, via plugins that integrate with specific cloud providers, and can manage nodes for you whilst optimizing for overall cost.

Karpenter automatically launches just the right compute resources to handle your cluster's applications. It is designed to let you take full advantage of the cloud with fast and simple compute provisioning for Kubernetes clusters.

The Karpenter tool is designed to integrate with a cloud provider that provides API-driven server management, and where the price information for available servers is also available via a web API.

For example, if you start some more Pods in your cluster, the Karpenter tool might buy a new node that is larger than one of the nodes you are already using, and then shut down an existing node once the new node is in service.

Cloud provider integrations

There are integrations available between Karpenter's core and the following cloud providers:

Descheduler

The descheduler can help you consolidate Pods onto a smaller number of nodes, to help with automatic scale down when the cluster has space capacity.

Sizing a workload based on cluster size

Cluster proportional autoscaler

For workloads that need to be scaled based on the size of the cluster (for example cluster-dns or other system components), you can use the Cluster Proportional Autoscaler.

The Cluster Proportional Autoscaler watches the number of schedulable nodes and cores, and scales the number of replicas of the target workload accordingly.

Cluster proportional vertical autoscaler

If the number of replicas should stay the same, you can scale your workloads vertically according to the cluster size using the Cluster Proportional Vertical Autoscaler. This project is in beta and can be found on GitHub.

While the Cluster Proportional Autoscaler scales the number of replicas of a workload, the Cluster Proportional Vertical Autoscaler adjusts the resource requests for a workload (for example a Deployment or DaemonSet) based on the number of nodes and/or cores in the cluster.

What's next

12 - Installing Addons

Add-ons extend the functionality of Kubernetes.

This page lists some of the available add-ons and links to their respective installation instructions. The list does not try to be exhaustive.

Networking and Network Policy

  • ACI provides integrated container networking and network security with Cisco ACI.
  • Antrea operates at Layer 3/4 to provide networking and security services for Kubernetes, leveraging Open vSwitch as the networking data plane. Antrea is a CNCF project at the Sandbox level.
  • Calico is a networking and network policy provider. Calico supports a flexible set of networking options so you can choose the most efficient option for your situation, including non-overlay and overlay networks, with or without BGP. Calico uses the same engine to enforce network policy for hosts, pods, and (if using Istio & Envoy) applications at the service mesh layer.
  • Canal unites Flannel and Calico, providing networking and network policy.
  • Cilium is a networking, observability, and security solution with an eBPF-based data plane. Cilium provides a simple flat Layer 3 network with the ability to span multiple clusters in either a native routing or overlay/encapsulation mode, and can enforce network policies on L3-L7 using an identity-based security model that is decoupled from network addressing. Cilium can act as a replacement for kube-proxy; it also offers additional, opt-in observability and security features. Cilium is a CNCF project at the Graduated level.
  • CNI-Genie enables Kubernetes to seamlessly connect to a choice of CNI plugins, such as Calico, Canal, Flannel, or Weave. CNI-Genie is a CNCF project at the Sandbox level.
  • Contiv provides configurable networking (native L3 using BGP, overlay using vxlan, classic L2, and Cisco-SDN/ACI) for various use cases and a rich policy framework. Contiv project is fully open sourced. The installer provides both kubeadm and non-kubeadm based installation options.
  • Contrail, based on Tungsten Fabric, is an open source, multi-cloud network virtualization and policy management platform. Contrail and Tungsten Fabric are integrated with orchestration systems such as Kubernetes, OpenShift, OpenStack and Mesos, and provide isolation modes for virtual machines, containers/pods and bare metal workloads.
  • Flannel is an overlay network provider that can be used with Kubernetes.
  • Gateway API is an open source project managed by the SIG Network community and provides an expressive, extensible, and role-oriented API for modeling service networking.
  • Knitter is a plugin to support multiple network interfaces in a Kubernetes pod.
  • Multus is a Multi plugin for multiple network support in Kubernetes to support all CNI plugins (e.g. Calico, Cilium, Contiv, Flannel), in addition to SRIOV, DPDK, OVS-DPDK and VPP based workloads in Kubernetes.
  • OVN-Kubernetes is a networking provider for Kubernetes based on OVN (Open Virtual Network), a virtual networking implementation that came out of the Open vSwitch (OVS) project. OVN-Kubernetes provides an overlay based networking implementation for Kubernetes, including an OVS based implementation of load balancing and network policy.
  • Nodus is an OVN based CNI controller plugin to provide cloud native based Service function chaining(SFC).
  • NSX-T Container Plug-in (NCP) provides integration between VMware NSX-T and container orchestrators such as Kubernetes, as well as integration between NSX-T and container-based CaaS/PaaS platforms such as Pivotal Container Service (PKS) and OpenShift.
  • Nuage is an SDN platform that provides policy-based networking between Kubernetes Pods and non-Kubernetes environments with visibility and security monitoring.
  • Romana is a Layer 3 networking solution for pod networks that also supports the NetworkPolicy API.
  • Spiderpool is an underlay and RDMA networking solution for Kubernetes. Spiderpool is supported on bare metal, virtual machines, and public cloud environments.
  • Weave Net provides networking and network policy, will carry on working on both sides of a network partition, and does not require an external database.

Service Discovery

  • CoreDNS is a flexible, extensible DNS server which can be installed as the in-cluster DNS for pods.

Visualization & Control

  • Dashboard is a dashboard web interface for Kubernetes.
  • Weave Scope is a tool for visualizing your containers, Pods, Services and more.

Infrastructure

Instrumentation

Legacy Add-ons

There are several other add-ons documented in the deprecated cluster/addons directory.

Well-maintained ones should be linked to here. PRs welcome!