Observability and debugging

A short order of checks

kubectl get pods shows a status word: Pending, CrashLoopBackOff, Error, Running, Completed. The word is not the diagnosis. kubectl describe pod puts events at the bottom: image pull errors, failed mounts, probe failures, and scheduling complaints.

kubectl logs prints what the process wrote. --previous is the last crashed container. If the process never started, the log is empty and the event is the useful part.

Debugging without a new screenshot tool

kubectl exec works when the container is running and includes a shell. Many minimal images do not. kubectl debug can attach an ephemeral container. That feature depends on the cluster version and on permission to create ephemeral containers. If the command is rejected, read the error. Do not invent a node SSH step as the default.

kubectl top needs metrics-server or an equivalent. A local kind cluster often does not have it. Missing metrics are not proof the Pod is idle.

Probes show up as events

A readiness probe that never passes leaves the Pod out of the Service. A liveness probe that is too aggressive restarts a slow process. Describe the Pod and read the probe events before you change the image.

Lab: break a Pod on purpose

Use an image tag that does not exist. Describe the Pod and find the pull error. Then delete the Pod.

kubectl run bad --image=nginx:this-tag-does-not-exist --restart=Never
kubectl describe pod bad | tail -n 20
kubectl delete pod bad --wait=false

Mark lesson complete

Concept summary

• Describe before you guess. Events are at the bottom.
• logs --previous is the crashed process. An empty log can mean it never started.
• kubectl top needs a metrics pipeline that local clusters often lack.

Knowledge check

Official sources

Checked against the linked documentation in October 2026. A later release can move a flag, a price, or a supported pair. Use the page for the version you installed.

• Debug Pods
• kubectl debug
• Events