k8s（Kubernetes 搜索）

Resource Management

• requests = guaranteed minimum — scheduler uses this for placement
• limits = maximum allowed — exceeding memory = OOMKilled, CPU = throttled
• No limits = can consume entire node — always set production limits
• requests without limits = burstable — can use more if available

Probes

• readinessProbe controls traffic — fails = removed from Service endpoints
• livenessProbe restarts container — fails = container killed and restarted
• startupProbe for slow starts — disables liveness/readiness until success
• Don't use same endpoint for liveness and readiness — liveness should be minimal health check

Probe Pitfalls

• Liveness probe checking dependencies — if DB down, all pods restart indefinitely
• initialDelaySeconds too short — pod killed before app starts
• timeoutSeconds too short — slow response = restart loop
• HTTP probe to HTTPS endpoint — needs scheme: HTTPS

Labels and Selectors

• Service selector must match Pod labels exactly — typo = no endpoints
• Deployment selector is immutable — can't change after creation
• Use consistent labeling scheme — app, version, environment
• matchExpressions for complex selection — In, NotIn, Exists

ConfigMaps and Secrets

• ConfigMap changes don't restart pods — mount as volume for auto-update, or restart manually
• Secrets are base64 encoded, not encrypted — use external secrets manager for sensitive data
• envFrom imports all keys — env.valueFrom for specific keys
• Volume mount makes files — subPath for single file without replacing directory

Networking

• ClusterIP internal only — default, only accessible within cluster
• NodePort exposes on node IP — 30000-32767 range, not for production
• LoadBalancer provisions cloud LB — works only in supported environments
• Ingress needs Ingress Controller — nginx-ingress, traefik, etc. installed separately

Persistent Storage

• PVC binds to PV — must match capacity and access modes
• storageClassName must match — or use "" for no dynamic provisioning
• ReadWriteOnce = single node — ReadWriteMany needed for multi-pod
• Pod deletion doesn't delete PVC — persistentVolumeReclaimPolicy controls PV fate

Common Mistakes

• kubectl apply vs create — apply for declarative (can update), create for imperative (fails if exists)
• Forgetting namespace — -n namespace or set context default
• Image tag latest in production — no version pinning, unpredictable updates
• Not setting imagePullPolicy — Always for latest tag, IfNotPresent for versioned
• Service port vs targetPort — port is Service's, targetPort is container's

Debugging

• kubectl describe pod for events — shows scheduling failures, probe failures
• kubectl logs -f pod for logs — -p for previous container (after crash)
• kubectl exec -it pod -- sh for shell — debug inside container
• kubectl get events --sort-by=.lastTimestamp — cluster-wide events timeline

RBAC

• ServiceAccount per workload — not default, for least privilege
• Role is namespaced — ClusterRole is cluster-wide
• RoleBinding binds Role to user/SA — ClusterRoleBinding for cluster-wide
• Check permissions: kubectl auth can-i verb resource --as=system:serviceaccount:ns:sa