Troubleshoot Running Pods

  • Status: Pending
  • Version: Alpha
  • Implementation Owner: @verb

This proposal seeks to add first class support for troubleshooting by creating a mechanism to execute a shell or other troubleshooting tools inside a running pod without requiring that the associated container images include such tools.



Many developers of native Kubernetes applications wish to treat Kubernetes as an execution platform for custom binaries produced by a build system. These users can forgo the scripted OS install of traditional Dockerfiles and instead COPY the output of their build system into a container image built FROM scratch or a distroless container image. This confers several advantages:

  1. Minimal images lower operational burden and reduce attack vectors.
  2. Immutable images improve correctness and reliability.
  3. Smaller image size reduces resource usage and speeds deployments.

The disadvantage of using containers built FROM scratch is the lack of system binaries provided by an Operating System image makes it difficult to troubleshoot running containers. Kubernetes should enable one to troubleshoot pods regardless of the contents of the container images.

Operations and Support

As Kubernetes gains in popularity, it’s becoming the case that a person troubleshooting an application is not necessarily the person who built it. Operations staff and Support organizations want the ability to attach a “known good” or automated debugging environment to a pod.


A solution to troubleshoot arbitrary container images MUST:

  • troubleshoot arbitrary running containers with minimal prior configuration
  • allow access to namespaces and the file systems of individual containers
  • fetch troubleshooting utilities at debug time rather than at the time of pod creation
  • be compatible with admission controllers and audit logging
  • allow discovery of debugging status
  • support arbitrary runtimes via the CRI (possibly with reduced feature set)
  • require no administrative access to the node
  • have an excellent user experience (i.e. should be a feature of the platform rather than config-time trickery)
  • have no inherent side effects to the running container image

Feature Summary

Any new debugging functionality will require training users. We can ease the transition by building on an existing usage pattern. We will create a new command, kubectl debug, which parallels an existing command, kubectl exec. Whereas kubectl exec runs a process in a container, kubectl debug will be similar but run a container in a pod.

A container created by kubectl debug is a Debug Container. Just like a process run by kubectl exec, a Debug Container is not part of the pod spec and has no resource stored in the API. Unlike kubectl exec, a Debug Container does have status that is reported in v1.PodStatus and displayed by kubectl describe pod.

For example, the following command would attach to a newly created container in a pod:

kubectl debug -c debug-shell --image=debian target-pod -- bash

It would be reasonable for Kubernetes to provide a default container name and image, making the minimal possible debug command:

kubectl debug target-pod

This creates an interactive shell in a pod which can examine and signal other processes in the pod. It has access to the same network and IPC as processes in the pod. It can access the filesystem of other processes by /proc/$PID/root. As is already the case with regular containers, Debug Containers can enter arbitrary namespaces of another container via nsenter when run with CAP_SYS_ADMIN.

Please see the User Stories section for additional examples and Alternatives Considered for the considerable list of other solutions we considered.

Implementation Details

The implementation of kubectl debug closely mirrors the implementation of kubectl exec, with most of the complexity implemented in the kubelet. How functionality like this best fits into Kubernetes API has been contentious. In order to make progress, we will start with the smallest possible API change, extending /exec to support Debug Containers, and iterate.

From the perspective of the user, there’s a new command, kubectl debug, that creates a Debug Container and attaches to its console. We believe a new command will be less confusing for users than overloading kubectl exec with a new concept. Users give Debug Containers a name (e.g. “debug” or “shell”) which can subsequently be used to reattach and is reported by kubectl describe.

Kubernetes API Changes

Chosen Solution: “exec++”

We will extend v1.Pod’s /exec subresource to support “executing” container images. The current /exec endpoint must implement GET to support streaming for all clients. We don’t want to encode a (potentially large) v1.Container as an HTTP parameter, so we must extend v1.PodExecOptions with the specific fields required for creating a Debug Container:

// PodExecOptions is the query options to a Pod's remote exec call
type PodExecOptions struct {
        // EphemeralContainerName is the name of an ephemeral container in which the
        // command ought to be run. Either both EphemeralContainerName and
        // EphemeralContainerImage fields must be set, or neither.
        EphemeralContainerName *string `json:"ephemeralContainerName,omitempty" ...`

        // EphemeralContainerImage is the image of an ephemeral container in which the command
        // ought to be run. Either both EphemeralContainerName and EphemeralContainerImage
        // fields must be set, or neither.
        EphemeralContainerImage *string `json:"ephemeralContainerImage,omitempty" ...`

After creating the Debug Container, the kubelet will upgrade the connection to streaming and perform an attach to the container’s console. If disconnected, the Debug Container can be reattached using the pod’s /attach endpoint with EphemeralContainerName.

Debug Containers cannot be removed via the API and instead the process must terminate. While not ideal, this parallels existing behavior of kubectl exec. To kill a Debug Container one would attach and exit the process interactively or create a new Debug Container to send a signal with kill(1) to the original process.

Alternative 1: Debug Subresource

Rather than extending an existing subresource, we could create a new, non-streaming debug subresource. We would create a new API Object:

// DebugContainer describes a container to attach to a running pod for troubleshooting.
type DebugContainer struct {

       // Name is the name of the Debug Container. Its presence will cause
        // exec to create a Debug Container rather than performing a runtime exec.
        Name string `json:"name,omitempty" ...`

        // Image is an optional container image name that will be used to for the Debug
        // Container in the specified Pod with Command as ENTRYPOINT. If omitted a
        // default image will be used.
        Image string `json:"image,omitempty" ...`

The pod would gain a new /debug subresource that allows the following:

  1. A POST of a PodDebugContainer to /api/v1/namespaces/$NS/pods/$POD_NAME/debug/$NAME to create Debug Container named $NAME running in pod $POD_NAME.
  2. A DELETE of /api/v1/namespaces/$NS/pods/$POD_NAME/debug/$NAME will stop the Debug Container $NAME in pod $POD_NAME.

Once created, a client would attach to the console of a debug container using the existing attach endpoint, /api/v1/namespaces/$NS/pods/$POD_NAME/attach.

However, this pattern does not resemble any other current usage of the API, so we prefer to start with exec++ and reevaluate if we discover a compelling reason.

Alternative 2: Declarative Configuration

Using subresources is an imperative style API where the client instructs the kubelet to perform an action, but in general Kubernetes prefers declarative APIs where the client declares a state for Kubernetes to enact.

We could implement this in a declarative manner by creating a new EphemeralContainer type:

type EphemeralContainer struct {

        Spec EphemeralContainerSpec
        Status v1.ContainerStatus

EphemeralContainerSpec is similar to v1.Container, but contains only fields relevant to Debug Containers:

type EphemeralContainerSpec struct {
        // Target is the pod in which to run the EphemeralContainer
        // Required.
        Target v1.ObjectReference

        Name string
        Image String
        Command []string
        Args []string
        ImagePullPolicy PullPolicy
        SecurityContext *SecurityContext

A new controller in the kubelet would watch for EphemeralContainers and create/delete debug containers. EphemeralContainer.Status would be updated by the kubelet at the same time it updates ContainerStatus for regular and init containers. Clients would create a new EphemeralContainer object, wait for it to be started and then attach using the pod’s attach subresource and the name of the EphemeralContainer.

Debugging is inherently imperative, however, rather than a state for Kubernetes to enforce. Once a Debug Container is started it should not be automatically restarted, for example. This solution imposes additionally complexity and dependencies on the kubelet, but it’s not yet clear if the complexity is justified.

Debug Container Status

The status of a Debug Container is reported in a new field in v1.PodStatus:

type PodStatus struct {
        EphemeralContainerStatuses []v1.ContainerStatus

This status is only populated for Debug Containers, but there’s interest in tracking status for traditional exec in a similar manner.

Note that Command and Args would have to be tracked in the status object because there is no spec for Debug Containers or exec. These must either be made available by the runtime or tracked by the kubelet. For Debug Containers this could be stored as runtime labels, but the kubelet currently has no method of storing state across restarts for exec. Solving this problem for exec is out of scope for Debug Containers, but we will look for a solution as we implement this feature.

EphemeralContainerStatuses is populated by the kubelet in the same way as regular and init container statuses. This is sent to the API server and displayed by kubectl describe pod.

Creating Debug Containers

  1. kubectl invokes the exec API as described in the preceding section.
  2. The API server checks for name collisions with existing containers, performs admission control and proxies the connection to the kubelet’s /exec/$NS/$POD_NAME/$CONTAINER_NAME endpoint.
  3. The kubelet instructs the Runtime Manager to create a Debug Container.
  4. The runtime manager uses the existing startContainer() method to create a container in an existing pod. startContainer() has one modification for Debug Containers: it creates a new runtime label (e.g. a docker label) that identifies this container as a Debug Container.
  5. After creating the container, the kubelet schedules an asynchronous update of PodStatus. The update publishes the debug container status to the API server at which point the Debug Container becomes visible via kubectl describe pod.
  6. The kubelet will upgrade the connection to streaming and attach to the container’s console.

Rather than performing the implicit attach the kubelet could return success to the client and require the client to perform an explicit attach, but the implicit attach maintains consistent semantics across /exec rather than varying behavior based on parameters.

The apiserver detects container name collisions with both containers in the pod spec and other running Debug Containers by checking EphemeralContainerStatuses. In a race to create two Debug Containers with the same name, the API server will pass both requests and the kubelet must return an error to all but one request.

There are no limits on the number of Debug Containers that can be created in a pod, but exceeding a pod’s resource allocation may cause the pod to be evicted.

Restarting and Reattaching Debug Containers

Debug Containers will never be restarted automatically. It is possible to replace a Debug Container that has exited by re-using a Debug Container name. It is an error to attempt to replace a Debug Container that is still running, which is detected by both the API server and the kubelet.

One can reattach to a Debug Container using kubectl attach. When supported by a runtime, multiple clients can attach to a single debug container and share the terminal. This is supported by Docker.

Killing Debug Containers

Debug containers will not be killed automatically until the pod (specifically, the pod sandbox) is destroyed. Debug Containers will stop when their command exits, such as exiting a shell. Unlike kubectl exec, processes in Debug Containers will not receive an EOF if their connection is interrupted.

Container Lifecycle Changes

Implementing debug requires no changes to the Container Runtime Interface as it’s the same operation as creating a regular container. The following changes are necessary in the kubelet:

  1. SyncPod() must not kill any Debug Container even though it is not part of the pod spec.
  2. As an exception to the above, SyncPod() will kill Debug Containers when the pod sandbox changes since a lone Debug Container in an abandoned sandbox is not useful. Debug Containers are not automatically started in the new sandbox.
  3. convertStatusToAPIStatus() must sort Debug Containers status into EphemeralContainerStatuses similar to as it does for InitContainerStatuses
  4. The kubelet must preserve ContainerStatus on debug containers for reporting.
  5. Debug Containers must be excluded from calculation of pod phase and condition

It’s worth noting some things that do not change:

  1. KillPod() already operates on all running containers returned by the runtime.
  2. Containers created prior to this feature being enabled will have a containerType of "". Since this does not match "EPHEMERAL" the special handling of Debug Containers is backwards compatible.

Security Considerations

Debug Containers have no additional privileges above what is available to any v1.Container. It’s the equivalent of configuring an shell container in a pod spec but created on demand.

Admission plugins that guard /exec must be updated for the new parameters. In particular, they should enforce the same container image policy on the Image parameter as is enforced for regular containers. During the alpha phase we will additionally support a container image whitelist as a kubelet flag to allow cluster administrators to easily constraint debug container images.

Additional Consideration

  1. Debug Containers are intended for interactive use and always have TTY and Stdin enabled.
  2. There are no guaranteed resources for ad-hoc troubleshooting. If troubleshooting causes a pod to exceed its resource limit it may be evicted.
  3. There’s an output stream race inherent to creating then attaching a container which causes output generated between the start and attach to go to the log rather than the client. This is not specific to Debug Containers and exists because Kubernetes has no mechanism to attach a container prior to starting it. This larger issue will not be addressed by Debug Containers, but Debug Containers would benefit from future improvements or work arounds.
  4. We do not want to describe Debug Containers using v1.Container. This is to reinforce that Debug Containers are not general purpose containers by limiting their configurability. Debug Containers should not be used to build services.
  5. Debug Containers are of limited usefulness without a shared PID namespace. If a pod is configured with isolated PID namespaces, the Debug Container will join the PID namespace of the target container. Debug Containers will not be available with runtimes that do not implement PID namespace sharing in some form.

Implementation Plan

Alpha Release

Goals and Non-Goals for Alpha Release

We’re targeting an alpha release in Kubernetes 1.9 that includes the following basic functionality:

  • Support in the kubelet for creating debug containers in a running pod
  • A kubectl debug command to initiate a debug container
  • kubectl describe pod will list status of debug containers running in a pod

Functionality will be hidden behind an alpha feature flag and disabled by default. The following are explicitly out of scope for the 1.9 alpha release:

  • Exited Debug Containers will be garbage collected as regular containers and may disappear from the list of Debug Container Statuses.
  • Security Context for the Debug Container is not configurable. It will always be run with CAP_SYS_PTRACE and CAP_SYS_ADMIN.
  • Image pull policy for the Debug Container is not configurable. It will always be run with PullAlways.

kubelet Implementation

Debug Containers are implemented in the kubelet’s generic runtime manager. Performing this operation with a legacy (non-CRI) runtime will result in a not implemented error. Implementation in the kubelet will be split into the following steps:

Step 1: Container Type

The first step is to add a feature gate to ensure all changes are off by default. This will be added in the pkg/features DefaultFeatureGate.

The runtime manager stores metadata about containers in the runtime via labels (e.g. docker labels). These labels are used to populate the fields of kubecontainer.ContainerStatus. Since the runtime manager needs to handle Debug Containers differently in a few situations, we must add a new piece of metadata to distinguish Debug Containers from regular containers.

startContainer() will be updated to write a new label io.kubernetes.container.type to the runtime. Existing containers will be started with a type of REGULAR or INIT. When added in a subsequent step, Debug Containers will start with the type EPHEMERAL.

Step 2: Creation and Handling of Debug Containers

This step adds methods for creating debug containers, but doesn’t yet modify the kubelet API. Since the runtime manager discards runtime (e.g. docker) labels after populating kubecontainer.ContainerStatus, the label value will be stored in a the new field ContainerStatus.Type so it can be used by SyncPod().

The kubelet gains a RunDebugContainer() method which accepts a v1.Container and passes it on to the Runtime Manager’s RunDebugContainer() if implemented. Currently only the Generic Runtime Manager (i.e. the CRI) implements the DebugContainerRunner interface.

The Generic Runtime Manager’s RunDebugContainer() calls startContainer() to create the Debug Container. Additionally, SyncPod() is modified to skip Debug Containers unless the sandbox is restarted.

Step 3: kubelet API changes

The kubelet exposes the new functionality in its existing /exec/ endpoint. ServeExec() constructs a v1.Container based on PodExecOptions, calls RunDebugContainer(), and performs the attach.

Step 4: Reporting EphemeralContainerStatus

The last major change to the kubelet is to populate v1.PodStatus.EphemeralContainerStatuses based on the kubecontainer.ContainerStatus for the Debug Container.

Kubernetes API Changes

There are two changes to be made to the Kubernetes, which will be made independently:

  1. v1.PodExecOptions must be extended with new fields.
  2. v1.PodStatus gains a new field to hold Debug Container statuses.

In all cases, new fields will be prepended with Alpha for the duration of this feature’s alpha status.

kubectl changes

In anticipation of this change, #46151 added a kubectl alpha command to contain alpha features. We will add kubectl alpha debug to invoke Debug Containers. kubectl does not use feature gates, so kubectl alpha debug will be visible by default in kubectl 1.9 and return an error when used on a cluster with the feature disabled.

kubectl describe pod will report the contents of EphemeralContainerStatuses when not empty as it means the feature is enabled. The field will be hidden when empty.


We’ve researched many options over the life of this proposal. These Appendices are included as optional reference material. It’s not necessary to read this material in order to understand the proposal in its current form.

Appendix 1: User Stories

These user stories are intended to give examples how this proposal addresses the above requirements.


Jonas runs a service “neato” that consists of a statically compiled Go binary running in a minimal container image. One of the its pods is suddenly having trouble connecting to an internal service. Being in operations, Jonas wants to be able to inspect the running pod without restarting it, but he doesn’t necessarily need to enter the container itself. He wants to:

  1. Inspect the filesystem of target container
  2. Execute debugging utilities not included in the container image
  3. Initiate network requests from the pod network namespace

This is achieved by running a new “debug” container in the pod namespaces. His troubleshooting session might resemble:

% kubectl debug -it -m debian neato-5thn0 -- bash
root@debug-image:~# ps x
    1 ?        Ss     0:00 /pause
   13 ?        Ss     0:00 bash
   26 ?        Ss+    0:00 /neato
  107 ?        R+     0:00 ps x
root@debug-image:~# cat /proc/26/root/etc/resolv.conf
search default.svc.cluster.local svc.cluster.local cluster.local
options ndots:5
root@debug-image:~# dig @ neato.svc.cluster.local.

; <<>> DiG 9.9.5-9+deb8u6-Debian <<>> @ neato.svc.cluster.local.
; (1 server found)
;; global options: +cmd
;; connection timed out; no servers could be reached

Thus Jonas discovers that the cluster’s DNS service isn’t responding.


Thurston is debugging a tricky issue that’s difficult to reproduce. He can’t reproduce the issue with the debug build, so he attaches a debug container to one of the pods exhibiting the problem:

% kubectl debug -it neato-5x9k3 -- sh
Defaulting container name to debug.
/ # ps x
    1 root       0:00 /pause
   13 root       0:00 /neato
   26 root       0:00 sh
   32 root       0:00 ps x
/ # gdb -p 13

He discovers that he needs access to the actual container, which he can achieve by installing busybox into the target container:

root@debug-image:~# cp /bin/busybox /proc/13/root
root@debug-image:~# nsenter -t 13 -m -u -p -n -r /busybox sh

BusyBox v1.22.1 (Debian 1:1.22.0-9+deb8u1) built-in shell (ash)
Enter 'help' for a list of built-in commands.

/ # ls -l /neato
-rwxr-xr-x    2 0        0           746888 May  4  2016 /neato

Note that running the commands referenced above require CAP_SYS_ADMIN and CAP_SYS_PTRACE.


Ginger is a security engineer tasked with running security audits across all of her company’s running containers. Even though his company has no standard base image, she’s able to audit all containers using:

% for pod in $(kubectl get -o name pod); do
    kubectl debug -m -p $pod /

Technical Support

Roy’s team provides support for his company’s multi-tenant cluster. He can access the Kubernetes API (as a viewer) on behalf of the users he’s supporting, but he does not have administrative access to nodes or a say in how the application image is constructed. When someone asks for help, Roy’s first step is to run his team’s autodiagnose script:

% kubectl debug nginx-pod-1234

Appendix 2: Requirements Analysis

Many people have proposed alternate solutions to this problem. This section discusses how the proposed solution meets all of the stated requirements and is intended to contrast the alternatives listed below.

Troubleshoot arbitrary running containers with minimal prior configuration. This solution requires no prior configuration.

Access to namespaces and the file systems of individual containers. This solution runs a container in the shared pod namespaces (e.g. network) and will attach to the PID namespace of a target container when not shared with the entire pod. It relies on the behavior of /proc/<pid>/root to provide access to filesystems of individual containers.

Fetch troubleshooting utilities at debug time. This solution uses normal container image distribution mechanisms to fetch images when the debug command is run.

Respect admission restrictions. Requests from kubectl are proxied through the apiserver and so are available to existing admission controllers. Plugins already exist to intercept exec and attach calls, but extending this to support debug has not yet been scoped.

Allow introspection of pod state using existing tools. The list of EphemeralContainerStatuses is never truncated. If a debug container has run in this pod it will appear here.

Support arbitrary runtimes via the CRI. This proposal is implemented entirely in the kubelet runtime manager and requires no changes in the individual runtimes.

Have an excellent user experience. This solution is conceptually straightforward and surfaced in a single kubectl command that “runs a thing in a pod”. Debug tools are distributed by container image, which is already well understood by users. There is no automatic copying of files or hidden paths.

By using container images, users are empowered to create custom debug images. Available images can be restricted by admission policy. Some examples of possible debug images:

  • A script that automatically gathers a debugging snapshot and uploads it to a cloud storage bucket before killing the pod.
  • An image with a shell modified to log every statement to an audit API.

Require no direct access to the node. This solution uses the standard streaming API.

Have no inherent side effects to the running container image. The target pod is not modified by default, but resources used by the debug container will be billed to the pod’s cgroup, which means it could be evicted. A future improvement could be to decrease the likelihood of eviction when there’s an active debug container.

Appendix 3: Alternatives Considered

Mutable Pod Spec

Rather than adding an operation to have Kubernetes attach a pod we could instead make the pod spec mutable so the client can generate an update adding a container. SyncPod() has no issues adding the container to the pod at that point, but an immutable pod spec has been a basic assumption in Kubernetes thus far and changing it carries risk. It’s preferable to keep the pod spec immutable as a best practice.

Ephemeral container

An earlier version of this proposal suggested running an ephemeral container in the pod namespaces. The container would not be added to the pod spec and would exist only as long as the process it ran. This has the advantage of behaving similarly to the current kubectl exec, but it is opaque and likely violates design assumptions. We could add constructs to track and report on both traditional exec process and exec containers, but this would probably be more work than adding to the pod spec. Both are generally useful, and neither precludes the other in the future, so we chose mutating the pod spec for expedience.

Attaching Container Type Volume

Combining container volumes (#831) with the ability to add volumes to the pod spec would get us most of the way there. One could mount a volume of debug utilities at debug time. Docker does not allow adding a volume to a running container, however, so this would require a container restart. A restart doesn’t meet our requirements for troubleshooting.

Rather than attaching the container at debug time, kubernetes could always attach a volume at a random path at run time, just in case it’s needed. Though this simplifies the solution by working within the existing constraints of kubectl exec, it has a sufficient list of minor limitations (detailed in #10834) to result in a poor user experience.

Inactive container

If Kubernetes supported the concept of an “inactive” container, we could configure it as part of a pod and activate it at debug time. In order to avoid coupling the debug tool versions with those of the running containers, we would need to ensure the debug image was pulled at debug time. The container could then be run with a TTY and attached using kubectl. We would need to figure out a solution that allows access the filesystem of other containers.

The downside of this approach is that it requires prior configuration. In addition to requiring prior consideration, it would increase boilerplate config. A requirement for prior configuration makes it feel like a workaround rather than a feature of the platform.

Implicit Empty Volume

Kubernetes could implicitly create an EmptyDir volume for every pod which would then be available as target for either the kubelet or a sidecar to extract a package of binaries.

Users would have to be responsible for hosting a package build and distribution infrastructure or rely on a public one. The complexity of this solution makes it undesirable.

Standalone Pod in Shared Namespace

Rather than inserting a new container into a pod namespace, Kubernetes could instead support creating a new pod with container namespaces shared with another, target pod. This would be a simpler change to the Kubernetes API, which would only need a new field in the pod spec to specify the target pod. To be useful, the containers in this “Debug Pod” should be run inside the namespaces (network, pid, etc) of the target pod but remain in a separate resource group (e.g. cgroup for container-based runtimes).

This would be a rather fundamental change to pod, which is currently treated as an atomic unit. The Container Runtime Interface has no provisions for sharing outside of a pod sandbox and would need a refactor. This could be a complicated change for non-container runtimes (e.g. hypervisor runtimes) which have more rigid boundaries between pods.

Effectively, Debug Pod must be implemented by the runtimes while Debug Containers are implemented by the kubelet. Minimizing change to the Kubernetes API is not worth the increased complexity for the kubelet and runtimes.

It could also be possible to implement a Debug Pod as a privileged pod that runs in the host namespace and interacts with the runtime directly to run a new container in the appropriate namespace. This solution would be runtime-specific and effectively pushes the complexity of debugging to the user. Additionally, requiring node-level access to debug a pod does not meet our requirements.

Exec from Node

The kubelet could support executing a troubleshooting binary from the node in the namespaces of the container. Once executed this binary would lose access to other binaries from the node, making it of limited utility and a confusing user experience.

This couples the debug tools with the lifecycle of the node, which is worse than coupling it with container images.