LEARN. PRACTICE. GROW.Concept and mechanism
Pending, ImagePullBackOff, and CrashLoopBackOff describe different situations. Start with state, events, and last termination to locate the failure phase. For a restarted container, current logs may omit the earlier cause; kubectl logs with --previous and the correct container and namespace may retrieve it when available. ImagePullBackOff requires reading the image retrieval error: reference, credentials, network, and registry can explain different causes. A denied event makes pull authorization a priority investigation. Avoid changing probes for an application whose container has not even been fetched.
Guided application
OOMKilled directs analysis toward memory, limits, and usage without alone proving a leak or insufficient sizing. Correlate load and release and assess a mitigation’s node impact. Liveness may restart containers; readiness indicates traffic eligibility; a startup probe supports handling slow startup before activating the other checks. For scheduling, requests must fit an eligible node. Two nodes with 700m free each cannot host a 1000m-request Pod by splitting it between them. Retain evidence and stability criteria after restoring service.
kubectl logs recon-7 -c app -n funds --previous queries the previous execution. Compare logs with kubectl describe pod recon-7 -n funds instead of repeatedly restarting without preserving the cause.
Common pitfalls
Confusing symptoms with cause; removing limits indiscriminately; expecting readiness to restart; summing node capacity for one Pod.
Related topics: Nodes and control plane · Cluster, context, and maintenance
Choose the next observation from the failure phase and recorded state.
Reference: Debug Pods · CKA Kubernetes v1.35