Miaoge Ge Coding Agent Skills Kubernetes Expert

Kubernetes Expert

Declarative, resilient, right-sized. Set resource requests/limits and correct probes, keep config out of images, and when debugging read the events first — they almost always name the cause.

When to Use

• Writing/reviewing manifests (Deployments, StatefulSets, Services, Ingress, Config/Secrets).
• Probes, resource requests/limits, autoscaling (HPA), rollouts.
• Debugging CrashLoopBackOff, ImagePullBackOff, OOMKilled, Pending, or networking.
• Templating with Helm/Kustomize.

When NOT to Use

• Building the container image → docker-expert.
• App-level bugs inside the container → relevant language skill.
• CI/CD pipeline wiring → github-master.

Core Principles

1. Pick the right workload

• Deployment for stateless apps, StatefulSet for stable identity/storage, DaemonSet for per-node, Job/CronJob for batch. Don't run databases as a plain Deployment.

2. Resources & scheduling

• Always set requests and limits. Requests drive scheduling and guarantees; limits cap usage. Missing requests → noisy-neighbor evictions and unschedulable surprises.
• Memory limit too low → OOMKilled. CPU limit throttles (doesn't kill). Size from real usage; set requests == limits for guaranteed/latency-sensitive pods.

3. Health & rollouts

• Readiness gates traffic (don't route until ready); liveness restarts a hung process; startup probe protects slow boots. A too-aggressive liveness probe causes restart loops — tune initialDelay/failureThreshold.
• RollingUpdate with sensible maxSurge/maxUnavailable; add PodDisruptionBudget for availability during drains. Pin image tags/digests, never :latest.

4. Config, secrets, security

• Config in ConfigMap, secrets in Secret (+ a real secrets manager / sealed-secrets for prod). Never bake them into images. Roll pods on config change (checksum annotation).
• Drop root, set securityContext (runAsNonRoot, read-only root FS), apply NetworkPolicy (default deny), and least-privilege RBAC.

5. Troubleshoot methodically

• kubectl get pods → describe pod (read Events) → logs (+ --previous for crashed) → get events --sort-by=.lastTimestamp. CrashLoop → check logs+probes; ImagePull → check tag/registry/secret; Pending → check resources/taints; OOMKilled → raise memory or fix the leak.

Common Mistakes

• No resource requests/limits → eviction, OOM, throttling, scheduling failures.
• Liveness probe doing heavy/slow checks → restart loops; use a cheap readiness check for traffic.
• :latest images → unpredictable rollouts, no rollback.
• Secrets in ConfigMaps or images → exposure.
• Editing live objects with kubectl edit instead of updating manifests → config drift; stay declarative/GitOps.
• One giant pod / multiple concerns per container → scale and fail independently instead.

Examples

Deployment with probes, resources, and security context

apiVersion: apps/v1
kind: Deployment
metadata: { name: web }
spec:
  replicas: 3
  selector: { matchLabels: { app: web } }
  template:
    metadata: { labels: { app: web } }
    spec:
      securityContext: { runAsNonRoot: true }
      containers:
        - name: web
          image: registry.example.com/web@sha256:…   # pinned by digest
          ports: [ { containerPort: 3000 } ]
          resources:
            requests: { cpu: "100m", memory: "128Mi" }
            limits:   { cpu: "500m", memory: "256Mi" }
          readinessProbe: { httpGet: { path: /health, port: 3000 }, initialDelaySeconds: 5 }
          livenessProbe:  { httpGet: { path: /health, port: 3000 }, periodSeconds: 10, failureThreshold: 6 }
          securityContext: { allowPrivilegeEscalation: false, readOnlyRootFilesystem: true }

See Also

• docker-expert — the images these workloads run.
• security-expert — RBAC, NetworkPolicy, and secret handling.
• performance-expert — right-sizing resources and HPA tuning.
• github-master — CD pipelines that apply manifests.

Source: Miaoge-Ge/coding-agent-skills — distributed by TomeVault.