In a Kubernetes interview, a strong troubleshooting answer starts with a diagnosis path: which command to run first, what to look for in its output, and which likely causes to investigate next. Knowing definitions is useful, but it does not explain how to isolate a failure. These eight common Kubernetes symptoms show where to start, how to interpret the findings, and how to explain the next diagnostic step.

Pending

Start with the pod description. Pending covers a pod that has not been scheduled, and can also include a pod waiting for its image to be pulled.

kubectl describe pod <pod> 

Read the Events at the bottom. A scheduler message such as 0/3 nodes are available: insufficient cpu means the pod's requests do not fit the available capacity on any node. Check requests against node allocatable resources. Other scheduling clues include untolerated taints, affinity rules that exclude nodes, and an unbound PVC. To see the newest cluster events in order, run:

kubectl get events --sort-by=.lastTimestamp 

An image pull error points to a different branch from a scheduler rejection, so use the event text to decide what to inspect next.

CrashLoopBackOff

Read the previous container's logs; the current container may have restarted and lost the output from the failed run.

kubectl logs <pod> --previous 

A missing configuration value or Secret can make the process exit at startup. So can a bad command or argument, or a dependency that is unreachable during initialisation. If the logs show slow startup followed by a liveness failure, check the probe settings. A startupProbe gives a slow application time to start before liveness checks begin. OOM is another possibility; inspect the termination reason and memory limit. CrashLoopBackOff describes Kubernetes waiting before retrying, not the cause of the crash. The delay starts at 10 seconds, doubles to 20 and 40 seconds, and caps at 300 seconds (5 minutes); it resets after the container runs for 10 minutes without problems.

ImagePullBackOff or ErrImagePull

Inspect the pod events to get the registry's actual response.

kubectl describe pod <pod> 

An event saying not found commonly points to a wrong image name or tag. unauthorized suggests a private registry with a missing imagePullSecret or expired credentials. A node that cannot reach the registry may instead have a network, DNS, proxy or TLS problem; registry rate limits can also block pulls. The Events line helps distinguish these cases. A pull that succeeds on your laptop does not prove that the cluster node can reach the same registry or use the same credentials. Fix the named image or tag, configure valid pull credentials, or restore the node's route to the registry, depending on the error.

OOMKilled or exit 137

Find the termination details, including the container's last state.

kubectl get pod <pod> -o yaml 

Look under lastState for reason: OOMKilled and the exit code. Exit 137 is SIGKILL (128+9). Paired with OOMKilled, it usually means the container exceeded its memory limit; if the node itself ran out of memory, the kernel can kill a container that was under its limit, so check node pressure too. Possible causes include a memory leak, a genuine load spike, or a limit below the application's working set. For example, a JVM heap larger than the container limit can leave too little room for other memory use. Check current usage with kubectl top pod <pod>; this requires metrics-server. If usage is consistently close to the limit, right-size the limit and application memory settings. If it spikes, investigate the workload or leak. CPU behaves differently: exceeding its limit throttles CPU, while memory over the limit can lead to a kill.

Service returns nothing

Check whether the Service has ready backend endpoints.

kubectl get endpointslices -n <namespace> -l kubernetes.io/service-name=<svc> 

An empty result means the Service has no endpoint slices with backends. Check that its selector matches the pod labels and that matching pods are ready. A failing readiness probe removes a pod from the Service's endpoints, even when the process is running. If slices exist, inspect their addresses and conditions to see which pods are included. This check separates a Service with no eligible backends from an application that has endpoints but still fails to respond.

Rollout stuck

Ask the Deployment whether its rollout is making progress.

kubectl rollout status deployment/<name> 

If progress times out, the Deployment condition can show Progressing=False with reason ProgressDeadlineExceeded. The new Pods may be failing readiness, unable to pull their image, or unable to schedule because the cluster lacks capacity for surge Pods. Run kubectl describe deployment <name> and inspect the new ReplicaSet's Pods and their Events to find which condition blocks progress. Once you know the cause, fix it or, if the release should be withdrawn, run kubectl rollout undo deployment/<name>. A rollback is a recovery action; the Deployment and Pod details explain why progress stopped.

Forbidden from the API

Repeat the failed permission check for the same identity, namespace, verb and resource.

kubectl auth can-i <verb> <resource> --as=<user> -n <namespace> 

A no means that identity lacks the requested permission in that namespace. Check for a missing RoleBinding, a binding to the wrong ServiceAccount or namespace, or a Role that omits the verb, resource or API group. For a ServiceAccount, impersonate it with --as=system:serviceaccount:<ns>:<name>. Compare the check with the exact request that received Forbidden, because changing the namespace or resource tests a different permission. RBAC grants are additive, so inspect the applicable bindings and roles to see which permission is absent.

Node NotReady

First find out whether one node or the whole cluster is affected.

kubectl get nodes -o wide 

Then inspect the affected node's conditions:

kubectl describe node <node> 

Check Ready, MemoryPressure, DiskPressure and PIDPressure. On the node, systemctl status kubelet and journalctl -u kubelet show kubelet health and recent errors. Check whether containerd is healthy, disk space is exhausted, kubelet certificates have expired, or the node can reach the API server. If one node is NotReady, focus on that machine and its route; if many changed together, investigate a shared dependency such as API server connectivity.

Say the path out loud

Use three steps: symptom, first command, hypothesis. For a Pending pod, you could say: “This pod is Pending, so I’d run kubectl describe pod checkout-7d9f and read its Events. If I see 0/3 nodes are available: insufficient cpu, the scheduler cannot fit its requests on the available nodes; I’d compare the requests with allocatable CPU and then consider adjusting requests or adding capacity.” This gives the interviewer a finding and a next check they can follow. For more practice, browse NostrobeAI's free Kubernetes interview questions with answers.