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

Return to the regular view of this page.


Technology for packaging an application along with its runtime dependencies.

Each container that you run is repeatable; the standardization from having dependencies included means that you get the same behavior wherever you run it.

Containers decouple applications from underlying host infrastructure. This makes deployment easier in different cloud or OS environments.

Container images

A container image is a ready-to-run software package, containing everything needed to run an application: the code and any runtime it requires, application and system libraries, and default values for any essential settings.

By design, a container is immutable: you cannot change the code of a container that is already running. If you have a containerized application and want to make changes, you need to build a new image that includes the change, then recreate the container to start from the updated image.

Container runtimes

The container runtime is the software that is responsible for running containers.

Kubernetes supports several container runtimes: Docker, containerd, CRI-O, and any implementation of the Kubernetes CRI (Container Runtime Interface).

What's next

1 - Images

A container image represents binary data that encapsulates an application and all its software dependencies. Container images are executable software bundles that can run standalone and that make very well defined assumptions about their runtime environment.

You typically create a container image of your application and push it to a registry before referring to it in a Pod

This page provides an outline of the container image concept.

Image names

Container images are usually given a name such as pause, example/mycontainer, or kube-apiserver. Images can also include a registry hostname; for example: fictional.registry.example/imagename, and possibly a port number as well; for example: fictional.registry.example:10443/imagename.

If you don't specify a registry hostname, Kubernetes assumes that you mean the Docker public registry.

After the image name part you can add a tag (as also using with commands such as docker and podman). Tags let you identify different versions of the same series of images.

Image tags consist of lowercase and uppercase letters, digits, underscores (_), periods (.), and dashes (-).
There are additional rules about where you can place the separator characters (_, -, and .) inside an image tag.
If you don't specify a tag, Kubernetes assumes you mean the tag latest.

Updating images

When you first create a Deployment, StatefulSet, Pod, or other object that includes a Pod template, then by default the pull policy of all containers in that pod will be set to IfNotPresent if it is not explicitly specified. This policy causes the kubelet to skip pulling an image if it already exists.

Image pull policy

The imagePullPolicy for a container and the tag of the image affect when the kubelet attempts to pull (download) the specified image.

Here's a list of the values you can set for imagePullPolicy and the effects these values have:

the image is pulled only if it is not already present locally.
every time the kubelet launches a container, the kubelet queries the container image registry to resolve the name to an image digest. If the kubelet has a container image with that exact digest cached locally, the kubelet uses its cached image; otherwise, the kubelet pulls the image with the resolved digest, and uses that image to launch the container.
the kubelet does not try fetching the image. If the image is somehow already present locally, the kubelet attempts to start the container; otherwise, startup fails. See pre-pulled images for more details.

The caching semantics of the underlying image provider make even imagePullPolicy: Always efficient, as long as the registry is reliably accessible. Your container runtime can notice that the image layers already exist on the node so that they don't need to be downloaded again.

To make sure the Pod always uses the same version of a container image, you can specify the image's digest; replace <image-name>:<tag> with <image-name>@<digest> (for example, image@sha256:45b23dee08af5e43a7fea6c4cf9c25ccf269ee113168c19722f87876677c5cb2).

When using image tags, if the image registry were to change the code that the tag on that image represents, you might end up with a mix of Pods running the old and new code. An image digest uniquely identifies a specific version of the image, so Kubernetes runs the same code every time it starts a container with that image name and digest specified. Specifying an image fixes the code that you run so that a change at the registry cannot lead to that mix of versions.

There are third-party admission controllers that mutate Pods (and pod templates) when they are created, so that the running workload is defined based on an image digest rather than a tag. That might be useful if you want to make sure that all your workload is running the same code no matter what tag changes happen at the registry.

Default image pull policy

When you (or a controller) submit a new Pod to the API server, your cluster sets the imagePullPolicy field when specific conditions are met:

  • if you omit the imagePullPolicy field, and the tag for the container image is :latest, imagePullPolicy is automatically set to Always;
  • if you omit the imagePullPolicy field, and you don't specify the tag for the container image, imagePullPolicy is automatically set to Always;
  • if you omit the imagePullPolicy field, and you specify the tag for the container image that isn't :latest, the imagePullPolicy is automatically set to IfNotPresent.

Required image pull

If you would like to always force a pull, you can do one of the following:

  • Set the imagePullPolicy of the container to Always.
  • Omit the imagePullPolicy and use :latest as the tag for the image to use; Kubernetes will set the policy to Always when you submit the Pod.
  • Omit the imagePullPolicy and the tag for the image to use; Kubernetes will set the policy to Always when you submit the Pod.
  • Enable the AlwaysPullImages admission controller.


When a kubelet starts creating containers for a Pod using a container runtime, it might be possible the container is in Waiting state because of ImagePullBackOff.

The status ImagePullBackOff means that a container could not start because Kubernetes could not pull a container image (for reasons such as invalid image name, or pulling from a private registry without imagePullSecret). The BackOff part indicates that Kubernetes will keep trying to pull the image, with an increasing back-off delay.

Kubernetes raises the delay between each attempt until it reaches a compiled-in limit, which is 300 seconds (5 minutes).

Multi-architecture images with image indexes

As well as providing binary images, a container registry can also serve a container image index. An image index can point to multiple image manifests for architecture-specific versions of a container. The idea is that you can have a name for an image (for example: pause, example/mycontainer, kube-apiserver) and allow different systems to fetch the right binary image for the machine architecture they are using.

Kubernetes itself typically names container images with a suffix -$(ARCH). For backward compatibility, please generate the older images with suffixes. The idea is to generate say pause image which has the manifest for all the arch(es) and say pause-amd64 which is backwards compatible for older configurations or YAML files which may have hard coded the images with suffixes.

Using a private registry

Private registries may require keys to read images from them.
Credentials can be provided in several ways:

  • Configuring Nodes to Authenticate to a Private Registry
    • all pods can read any configured private registries
    • requires node configuration by cluster administrator
  • Pre-pulled Images
    • all pods can use any images cached on a node
    • requires root access to all nodes to setup
  • Specifying ImagePullSecrets on a Pod
    • only pods which provide own keys can access the private registry
  • Vendor-specific or local extensions
    • if you're using a custom node configuration, you (or your cloud provider) can implement your mechanism for authenticating the node to the container registry.

These options are explained in more detail below.

Configuring nodes to authenticate to a private registry

If you run Docker on your nodes, you can configure the Docker container runtime to authenticate to a private container registry.

This approach is suitable if you can control node configuration.

Docker stores keys for private registries in the $HOME/.dockercfg or $HOME/.docker/config.json file. If you put the same file in the search paths list below, kubelet uses it as the credential provider when pulling images.

  • {--root-dir:-/var/lib/kubelet}/config.json
  • {cwd of kubelet}/config.json
  • ${HOME}/.docker/config.json
  • /.docker/config.json
  • {--root-dir:-/var/lib/kubelet}/.dockercfg
  • {cwd of kubelet}/.dockercfg
  • ${HOME}/.dockercfg
  • /.dockercfg

Here are the recommended steps to configuring your nodes to use a private registry. In this example, run these on your desktop/laptop:

  1. Run docker login [server] for each set of credentials you want to use. This updates $HOME/.docker/config.json on your PC.
  2. View $HOME/.docker/config.json in an editor to ensure it contains only the credentials you want to use.
  3. Get a list of your nodes; for example:
    • if you want the names: nodes=$( kubectl get nodes -o jsonpath='{range.items[*].metadata}{.name} {end}' )
    • if you want to get the IP addresses: nodes=$( kubectl get nodes -o jsonpath='{range .items[*].status.addresses[?(@.type=="ExternalIP")]}{.address} {end}' )
  4. Copy your local .docker/config.json to one of the search paths list above.
    • for example, to test this out: for n in $nodes; do scp ~/.docker/config.json root@"$n":/var/lib/kubelet/config.json; done

Verify by creating a Pod that uses a private image; for example:

kubectl apply -f - <<EOF
apiVersion: v1
kind: Pod
  name: private-image-test-1
    - name: uses-private-image
      image: $PRIVATE_IMAGE_NAME
      imagePullPolicy: Always
      command: [ "echo", "SUCCESS" ]
pod/private-image-test-1 created

If everything is working, then, after a few moments, you can run:

kubectl logs private-image-test-1

and see that the command outputs:


If you suspect that the command failed, you can run:

kubectl describe pods/private-image-test-1 | grep 'Failed'

In case of failure, the output is similar to:

  Fri, 26 Jun 2015 15:36:13 -0700    Fri, 26 Jun 2015 15:39:13 -0700    19    {kubelet node-i2hq}    spec.containers{uses-private-image}    failed        Failed to pull image "user/privaterepo:v1": Error: image user/privaterepo:v1 not found

You must ensure all nodes in the cluster have the same .docker/config.json. Otherwise, pods will run on some nodes and fail to run on others. For example, if you use node autoscaling, then each instance template needs to include the .docker/config.json or mount a drive that contains it.

All pods will have read access to images in any private registry once private registry keys are added to the .docker/config.json.

Interpretation of config.json

The interpretation of config.json varies between the original Docker implementation and the Kubernetes interpretation. In Docker, the auths keys can only specify root URLs, whereas Kubernetes allows glob URLs as well as prefix-matched paths. This means that a config.json like this is valid:

    "auths": {
        "*": {
            "auth": "…"

The root URL (* is matched by using the following syntax:

    { term }

    '*'         matches any sequence of non-Separator characters
    '?'         matches any single non-Separator character
    '[' [ '^' ] { character-range } ']'
                character class (must be non-empty)
    c           matches character c (c != '*', '?', '\\', '[')
    '\\' c      matches character c

    c           matches character c (c != '\\', '-', ']')
    '\\' c      matches character c
    lo '-' hi   matches character c for lo <= c <= hi

Image pull operations would now pass the credentials to the CRI container runtime for every valid pattern. For example the following container image names would match successfully:


The kubelet performs image pulls sequentially for every found credential. This means, that multiple entries in config.json are possible, too:

    "auths": {
        "": {
            "auth": "…"
        "": {
            "auth": "…"

If now a container specifies an image to be pulled, then the kubelet will try to download them from both authentication sources if one of them fails.

Pre-pulled images

By default, the kubelet tries to pull each image from the specified registry. However, if the imagePullPolicy property of the container is set to IfNotPresent or Never, then a local image is used (preferentially or exclusively, respectively).

If you want to rely on pre-pulled images as a substitute for registry authentication, you must ensure all nodes in the cluster have the same pre-pulled images.

This can be used to preload certain images for speed or as an alternative to authenticating to a private registry.

All pods will have read access to any pre-pulled images.

Specifying imagePullSecrets on a Pod

Kubernetes supports specifying container image registry keys on a Pod.

Creating a Secret with a Docker config

Run the following command, substituting the appropriate uppercase values:

kubectl create secret docker-registry <name> --docker-server=DOCKER_REGISTRY_SERVER --docker-username=DOCKER_USER --docker-password=DOCKER_PASSWORD --docker-email=DOCKER_EMAIL

If you already have a Docker credentials file then, rather than using the above command, you can import the credentials file as a Kubernetes Secrets.
Create a Secret based on existing Docker credentials explains how to set this up.

This is particularly useful if you are using multiple private container registries, as kubectl create secret docker-registry creates a Secret that only works with a single private registry.

Referring to an imagePullSecrets on a Pod

Now, you can create pods which reference that secret by adding an imagePullSecrets section to a Pod definition.

For example:

cat <<EOF > pod.yaml
apiVersion: v1
kind: Pod
  name: foo
  namespace: awesomeapps
    - name: foo
      image: janedoe/awesomeapp:v1
    - name: myregistrykey

cat <<EOF >> ./kustomization.yaml
- pod.yaml

This needs to be done for each pod that is using a private registry.

However, setting of this field can be automated by setting the imagePullSecrets in a ServiceAccount resource.

Check Add ImagePullSecrets to a Service Account for detailed instructions.

You can use this in conjunction with a per-node .docker/config.json. The credentials will be merged.

Use cases

There are a number of solutions for configuring private registries. Here are some common use cases and suggested solutions.

  1. Cluster running only non-proprietary (e.g. open-source) images. No need to hide images.
    • Use public images on the Docker hub.
      • No configuration required.
      • Some cloud providers automatically cache or mirror public images, which improves availability and reduces the time to pull images.
  2. Cluster running some proprietary images which should be hidden to those outside the company, but visible to all cluster users.
    • Use a hosted private Docker registry.
      • It may be hosted on the Docker Hub, or elsewhere.
      • Manually configure .docker/config.json on each node as described above.
    • Or, run an internal private registry behind your firewall with open read access.
      • No Kubernetes configuration is required.
    • Use a hosted container image registry service that controls image access
      • It will work better with cluster autoscaling than manual node configuration.
    • Or, on a cluster where changing the node configuration is inconvenient, use imagePullSecrets.
  3. Cluster with proprietary images, a few of which require stricter access control.
    • Ensure AlwaysPullImages admission controller is active. Otherwise, all Pods potentially have access to all images.
    • Move sensitive data into a "Secret" resource, instead of packaging it in an image.
  4. A multi-tenant cluster where each tenant needs own private registry.
    • Ensure AlwaysPullImages admission controller is active. Otherwise, all Pods of all tenants potentially have access to all images.
    • Run a private registry with authorization required.
    • Generate registry credential for each tenant, put into secret, and populate secret to each tenant namespace.
    • The tenant adds that secret to imagePullSecrets of each namespace.

If you need access to multiple registries, you can create one secret for each registry. Kubelet will merge any imagePullSecrets into a single virtual .docker/config.json

What's next

2 - Container Environment

This page describes the resources available to Containers in the Container environment.

Container environment

The Kubernetes Container environment provides several important resources to Containers:

  • A filesystem, which is a combination of an image and one or more volumes.
  • Information about the Container itself.
  • Information about other objects in the cluster.

Container information

The hostname of a Container is the name of the Pod in which the Container is running. It is available through the hostname command or the gethostname function call in libc.

The Pod name and namespace are available as environment variables through the downward API.

User defined environment variables from the Pod definition are also available to the Container, as are any environment variables specified statically in the Docker image.

Cluster information

A list of all services that were running when a Container was created is available to that Container as environment variables. This list is limited to services within the same namespace as the new Container's Pod and Kubernetes control plane services. Those environment variables match the syntax of Docker links.

For a service named foo that maps to a Container named bar, the following variables are defined:

FOO_SERVICE_HOST=<the host the service is running on>
FOO_SERVICE_PORT=<the port the service is running on>

Services have dedicated IP addresses and are available to the Container via DNS, if DNS addon is enabled. 

What's next

3 - Runtime Class

FEATURE STATE: Kubernetes v1.20 [stable]

This page describes the RuntimeClass resource and runtime selection mechanism.

RuntimeClass is a feature for selecting the container runtime configuration. The container runtime configuration is used to run a Pod's containers.


You can set a different RuntimeClass between different Pods to provide a balance of performance versus security. For example, if part of your workload deserves a high level of information security assurance, you might choose to schedule those Pods so that they run in a container runtime that uses hardware virtualization. You'd then benefit from the extra isolation of the alternative runtime, at the expense of some additional overhead.

You can also use RuntimeClass to run different Pods with the same container runtime but with different settings.


  1. Configure the CRI implementation on nodes (runtime dependent)
  2. Create the corresponding RuntimeClass resources

1. Configure the CRI implementation on nodes

The configurations available through RuntimeClass are Container Runtime Interface (CRI) implementation dependent. See the corresponding documentation (below) for your CRI implementation for how to configure.

The configurations have a corresponding handler name, referenced by the RuntimeClass. The handler must be a valid DNS label name.

2. Create the corresponding RuntimeClass resources

The configurations setup in step 1 should each have an associated handler name, which identifies the configuration. For each handler, create a corresponding RuntimeClass object.

The RuntimeClass resource currently only has 2 significant fields: the RuntimeClass name ( and the handler (handler). The object definition looks like this:

apiVersion:  # RuntimeClass is defined in the API group
kind: RuntimeClass
  name: myclass  # The name the RuntimeClass will be referenced by
  # RuntimeClass is a non-namespaced resource
handler: myconfiguration  # The name of the corresponding CRI configuration

The name of a RuntimeClass object must be a valid DNS subdomain name.


Once RuntimeClasses are configured for the cluster, using them is very simple. Specify a runtimeClassName in the Pod spec. For example:

apiVersion: v1
kind: Pod
  name: mypod
  runtimeClassName: myclass
  # ...

This will instruct the kubelet to use the named RuntimeClass to run this pod. If the named RuntimeClass does not exist, or the CRI cannot run the corresponding handler, the pod will enter the Failed terminal phase. Look for a corresponding event for an error message.

If no runtimeClassName is specified, the default RuntimeHandler will be used, which is equivalent to the behavior when the RuntimeClass feature is disabled.

CRI Configuration

For more details on setting up CRI runtimes, see CRI installation.


RuntimeClasses with dockershim must set the runtime handler to docker. Dockershim does not support custom configurable runtime handlers.


Runtime handlers are configured through containerd's configuration at /etc/containerd/config.toml. Valid handlers are configured under the runtimes section:


See containerd's config documentation for more details:


Runtime handlers are configured through CRI-O's configuration at /etc/crio/crio.conf. Valid handlers are configured under the crio.runtime table:

  runtime_path = "${PATH_TO_BINARY}"

See CRI-O's config documentation for more details.


FEATURE STATE: Kubernetes v1.16 [beta]

By specifying the scheduling field for a RuntimeClass, you can set constraints to ensure that Pods running with this RuntimeClass are scheduled to nodes that support it. If scheduling is not set, this RuntimeClass is assumed to be supported by all nodes.

To ensure pods land on nodes supporting a specific RuntimeClass, that set of nodes should have a common label which is then selected by the runtimeclass.scheduling.nodeSelector field. The RuntimeClass's nodeSelector is merged with the pod's nodeSelector in admission, effectively taking the intersection of the set of nodes selected by each. If there is a conflict, the pod will be rejected.

If the supported nodes are tainted to prevent other RuntimeClass pods from running on the node, you can add tolerations to the RuntimeClass. As with the nodeSelector, the tolerations are merged with the pod's tolerations in admission, effectively taking the union of the set of nodes tolerated by each.

To learn more about configuring the node selector and tolerations, see Assigning Pods to Nodes.

Pod Overhead

FEATURE STATE: Kubernetes v1.18 [beta]

You can specify overhead resources that are associated with running a Pod. Declaring overhead allows the cluster (including the scheduler) to account for it when making decisions about Pods and resources. To use Pod overhead, you must have the PodOverhead feature gate enabled (it is on by default).

Pod overhead is defined in RuntimeClass through the overhead fields. Through the use of these fields, you can specify the overhead of running pods utilizing this RuntimeClass and ensure these overheads are accounted for in Kubernetes.

What's next

4 - Container Lifecycle Hooks

This page describes how kubelet managed Containers can use the Container lifecycle hook framework to run code triggered by events during their management lifecycle.


Analogous to many programming language frameworks that have component lifecycle hooks, such as Angular, Kubernetes provides Containers with lifecycle hooks. The hooks enable Containers to be aware of events in their management lifecycle and run code implemented in a handler when the corresponding lifecycle hook is executed.

Container hooks

There are two hooks that are exposed to Containers:


This hook is executed immediately after a container is created. However, there is no guarantee that the hook will execute before the container ENTRYPOINT. No parameters are passed to the handler.


This hook is called immediately before a container is terminated due to an API request or management event such as a liveness/startup probe failure, preemption, resource contention and others. A call to the PreStop hook fails if the container is already in a terminated or completed state and the hook must complete before the TERM signal to stop the container can be sent. The Pod's termination grace period countdown begins before the PreStop hook is executed, so regardless of the outcome of the handler, the container will eventually terminate within the Pod's termination grace period. No parameters are passed to the handler.

A more detailed description of the termination behavior can be found in Termination of Pods.

Hook handler implementations

Containers can access a hook by implementing and registering a handler for that hook. There are two types of hook handlers that can be implemented for Containers:

  • Exec - Executes a specific command, such as, inside the cgroups and namespaces of the Container. Resources consumed by the command are counted against the Container.
  • HTTP - Executes an HTTP request against a specific endpoint on the Container.

Hook handler execution

When a Container lifecycle management hook is called, the Kubernetes management system executes the handler according to the hook action, httpGet and tcpSocket are executed by the kubelet process, and exec is executed in the container.

Hook handler calls are synchronous within the context of the Pod containing the Container. This means that for a PostStart hook, the Container ENTRYPOINT and hook fire asynchronously. However, if the hook takes too long to run or hangs, the Container cannot reach a running state.

PreStop hooks are not executed asynchronously from the signal to stop the Container; the hook must complete its execution before the TERM signal can be sent. If a PreStop hook hangs during execution, the Pod's phase will be Terminating and remain there until the Pod is killed after its terminationGracePeriodSeconds expires. This grace period applies to the total time it takes for both the PreStop hook to execute and for the Container to stop normally. If, for example, terminationGracePeriodSeconds is 60, and the hook takes 55 seconds to complete, and the Container takes 10 seconds to stop normally after receiving the signal, then the Container will be killed before it can stop normally, since terminationGracePeriodSeconds is less than the total time (55+10) it takes for these two things to happen.

If either a PostStart or PreStop hook fails, it kills the Container.

Users should make their hook handlers as lightweight as possible. There are cases, however, when long running commands make sense, such as when saving state prior to stopping a Container.

Hook delivery guarantees

Hook delivery is intended to be at least once, which means that a hook may be called multiple times for any given event, such as for PostStart or PreStop. It is up to the hook implementation to handle this correctly.

Generally, only single deliveries are made. If, for example, an HTTP hook receiver is down and is unable to take traffic, there is no attempt to resend. In some rare cases, however, double delivery may occur. For instance, if a kubelet restarts in the middle of sending a hook, the hook might be resent after the kubelet comes back up.

Debugging Hook handlers

The logs for a Hook handler are not exposed in Pod events. If a handler fails for some reason, it broadcasts an event. For PostStart, this is the FailedPostStartHook event, and for PreStop, this is the FailedPreStopHook event. You can see these events by running kubectl describe pod <pod_name>. Here is some example output of events from running this command:

  FirstSeen  LastSeen  Count  From                                                   SubObjectPath          Type      Reason               Message
  ---------  --------  -----  ----                                                   -------------          --------  ------               -------
  1m         1m        1      {default-scheduler }                                                          Normal    Scheduled            Successfully assigned test-1730497541-cq1d2 to gke-test-cluster-default-pool-a07e5d30-siqd
  1m         1m        1      {kubelet gke-test-cluster-default-pool-a07e5d30-siqd}  spec.containers{main}  Normal    Pulling              pulling image "test:1.0"
  1m         1m        1      {kubelet gke-test-cluster-default-pool-a07e5d30-siqd}  spec.containers{main}  Normal    Created              Created container with docker id 5c6a256a2567; Security:[seccomp=unconfined]
  1m         1m        1      {kubelet gke-test-cluster-default-pool-a07e5d30-siqd}  spec.containers{main}  Normal    Pulled               Successfully pulled image "test:1.0"
  1m         1m        1      {kubelet gke-test-cluster-default-pool-a07e5d30-siqd}  spec.containers{main}  Normal    Started              Started container with docker id 5c6a256a2567
  38s        38s       1      {kubelet gke-test-cluster-default-pool-a07e5d30-siqd}  spec.containers{main}  Normal    Killing              Killing container with docker id 5c6a256a2567: PostStart handler: Error executing in Docker Container: 1
  37s        37s       1      {kubelet gke-test-cluster-default-pool-a07e5d30-siqd}  spec.containers{main}  Normal    Killing              Killing container with docker id 8df9fdfd7054: PostStart handler: Error executing in Docker Container: 1
  38s        37s       2      {kubelet gke-test-cluster-default-pool-a07e5d30-siqd}                         Warning   FailedSync           Error syncing pod, skipping: failed to "StartContainer" for "main" with RunContainerError: "PostStart handler: Error executing in Docker Container: 1"
  1m         22s       2      {kubelet gke-test-cluster-default-pool-a07e5d30-siqd}  spec.containers{main}  Warning   FailedPostStartHook

What's next