CrashLoopBackOff, ImagePullBackOff and other Kubernetes pod errors
Kubernetes pod errors explained: CrashLoopBackOff, ImagePullBackOff, Pending, OOMKilled and more, with the cause, the check and the fix for each.
CrashLoopBackOff means a container in your pod starts, exits, and is being restarted by the kubelet with a growing delay between attempts. ImagePullBackOff means the node cannot download the container image and is retrying with a growing delay. The first is a problem inside the running application or its configuration; the second happens before the application ever starts.
This reference lists the common pod and container states, what usually causes each, and the first commands to run. Commands assume you have kubectl access to the namespace; replace the placeholders with your own names. Behaviour described below follows the official Kubernetes documentation as of October 2026.
Start with the right command
Most pod errors reveal themselves in three places. Check them in this order:
- Pod details and events: kubectl describe pod POD. The Events section at the bottom usually names the failure, and the container section shows the state, reason and exit code.
- Logs of the previous run: kubectl logs POD -c CONTAINER --previous. A crashing container has already been replaced, so the current logs may be empty.
- Namespace events in time order: kubectl get events --sort-by=.lastTimestamp. Useful when several pods fail together.
If the image has no shell, kubectl debug can attach an ephemeral debugging container to a running pod, or create a copy of a failing pod with a different command or image. The official Debug Pods page covers the basic workflow.
Pod phases versus container states
Two different things get confused. The pod phase is a coarse summary: Pending, Running, Succeeded, Failed or Unknown. The pod lifecycle documentation defines Pending as accepted but with containers not yet set up, which includes waiting for scheduling and downloading images.
Names like CrashLoopBackOff, ImagePullBackOff and ContainerCreating are not phases. They are reasons shown in the STATUS column of kubectl get pods, taken from a container being in the Waiting or Terminated state. A pod can be in phase Running while one of its containers is in CrashLoopBackOff.
Pending: the pod is not placed or not ready to start
Unschedulable
If kubectl describe shows a FailedScheduling event, the scheduler found no suitable node. Typical messages mention insufficient CPU or memory, node taints the pod does not tolerate, or a node selector or affinity rule nothing satisfies. Fixes: lower the resource requests, add node capacity, or relax the placement rules.
Waiting on a volume
A pod that references a PersistentVolumeClaim stays Pending while the claim is unbound. Run kubectl get pvc and kubectl describe pvc NAME: a missing storage class, no matching volume, or a zone mismatch are the usual reasons.
ContainerCreating
The pod is on a node and the kubelet is preparing it: pulling the image, mounting volumes, applying Secret data. A short stay is normal. A long one points to a slow pull, a volume that will not attach or mount, or a Secret or ConfigMap volume whose source does not exist. The events name which.
ErrImagePull and ImagePullBackOff
These are two stages of one problem. ErrImagePull appears on a failed attempt; after repeated failures the status switches to ImagePullBackOff while the kubelet waits before trying again. The Kubernetes images documentation says the delay keeps increasing until it reaches a compiled-in limit of 300 seconds (5 minutes). So the two statuses differ in timing, not in cause. The event message shows the real error.
Common causes, with the check for each:
- Wrong image name or tag. A typo, a tag that was never pushed, or a deleted tag. Check the exact string in kubectl describe pod and try pulling it from a machine with registry access.
- Private registry without credentials. The documentation names this as a cause. Create a Secret of type kubernetes.io/dockerconfigjson and reference it in imagePullSecrets on the pod or its ServiceAccount. The Secret must be in the same namespace as the pod, so the step is needed once per namespace.
- Registry rate limits or outage. The event text shows an HTTP error or timeout rather than "not found". Authenticate to the registry, mirror the image into your own registry, or retry later.
- Network or DNS problems on the node. The node cannot reach the registry host; check proxy settings and egress rules.
- Architecture mismatch. A registry can serve an image index that points to per-architecture manifests. If the image has no manifest for the node CPU (for example arm64 nodes and an amd64-only image), the pull fails or the container later exits with an exec format error.
- imagePullPolicy: Never. With this policy the kubelet never fetches the image, so startup fails if it is not already on the node.
The documentation also advises against the :latest tag in production and suggests pinning an image by digest, which guarantees the same content is pulled every time.
CrashLoopBackOff
The container ran and terminated, and Kubernetes keeps restarting it under the pod's restart policy. After exits, the kubelet restarts containers with an exponential delay of 10s, 20s, 40s and so on, capped at 300 seconds. It resets the timer once a container has run for 10 minutes without problems. That is why a crashing pod looks healthy for a few seconds, then fails again, and the restart counter climbs.
The status is a symptom, not a diagnosis. Work from the exit code and the last log lines:
- Run kubectl describe pod POD and read Last State: Terminated, with its reason and exit code.
- Run kubectl logs POD --previous for the output of the crashed run.
- Match the cause: an application error or bad startup command (exit code 1 or 127), a missing environment variable or config file, a dependency such as a database that is unreachable, a failing liveness probe that kills a healthy but slow container, or an out-of-memory kill.
Our guide to container exit codes explains what 137, 139 and 143 mean. If logs are empty because the process dies instantly, setting terminationMessagePolicy to FallbackToLogsOnError copies the last log chunk into the container status, as described in the Kubernetes task page Determine the Reason for Pod Failure.
Changes to the back-off in newer versions
The 300-second cap is no longer fixed everywhere. According to the lifecycle documentation, the KubeletCrashLoopBackOffMax feature gate has been beta and on by default since Kubernetes 1.35. It lets you set a per-node maximum restart delay between 1s and 300s in the kubelet configuration (crashLoopBackOff.maxContainerRestartPeriod). Separately, an alpha gate called ReduceDefaultCrashLoopBackOffDecay, off by default since 1.33, starts retries at 1s and caps them at 60s. Managed clusters may differ. If you are still planning upgrades, see our note on Kubernetes 1.34 end of life.
CreateContainerConfigError and RunContainerError
CreateContainerConfigError means the kubelet could not build the container configuration. The most common reason is a reference to a ConfigMap or Secret, or to a key inside one, that does not exist in the pod's namespace. The event names the missing object. Create it or fix the name or key.
RunContainerError means the container runtime failed to start the container after it was created. The event message carries the runtime error, such as an executable that is not found or a permission problem on a mounted path. Compare the pod's command and args with what the image actually contains.
OOMKilled
When a container exceeds its memory limit, the kernel kills it and the container's last state shows reason OOMKilled, usually with exit code 137. The pod then restarts and often lands in CrashLoopBackOff. Exit code 137 can also come from other kill signals, so confirm the reason field. Memory accounting differs between cgroup versions, so our piece on checking cgroup v1 versus v2 is worth a look on mixed node pools.
Evicted
The kubelet evicts pods when a node runs short of resources such as memory, disk or inodes. An evicted pod ends in phase Failed with reason Evicted, and the message says which resource was low. The lifecycle documentation states that a pod will not survive an eviction caused by lack of resources or node maintenance. A controller such as a Deployment creates a replacement. Look at node pressure with kubectl describe node NODE, set realistic requests and limits, and delete leftover Evicted pods once you have read the cause.
Stuck in Terminating
A deletion was requested, but the API object is still there. The Debug Pods page points to a finalizer combined with an admission webhook that blocks the update needed to remove the pod. Check for validating or mutating webhook configurations that target UPDATE on pods and fix or disable the faulty one.
Completed
Completed is not an error: it is the status of a pod whose containers all exited with code 0, typical for Jobs and one-off tasks.
More platform troubleshooting and release news is in our DevOps platforms section and DevOps tooling coverage; managed-cluster topics sit under Cloud.
Frequently asked questions
What is the difference between ImagePullBackOff and CrashLoopBackOff?
ImagePullBackOff means Kubernetes cannot download the image, so the container never starts. CrashLoopBackOff means the container did start but keeps exiting and being restarted with a growing delay.
What is the difference between ErrImagePull and ImagePullBackOff?
ErrImagePull is shown when a pull attempt fails; ImagePullBackOff is shown while the kubelet waits before the next attempt. Both have the same underlying causes, and the pod events contain the actual error message.
How long does CrashLoopBackOff last?
By default the restart delay starts at 10 seconds, doubles after each crash and is capped at 5 minutes, then resets after the container runs for 10 minutes without problems. The pod stays in this state until the underlying fault is fixed.
How do I fix ImagePullBackOff?
Read the error in kubectl describe pod, then check the image name and tag, add an imagePullSecret in the pod's namespace if the registry is private, and confirm the node can reach the registry. The next retry then succeeds, or you can delete the pod to force a new attempt.
What causes CreateContainerConfigError?
Usually a ConfigMap or Secret, or a key inside it, that the pod references does not exist in the same namespace. The pod events name the missing object.