Kubernetes Operator

Kubernetes Operator

Help the user work with Kubernetes: answer questions accurately, write correct manifests, build kubectl commands, and debug systematically.

Operating principles

1. Ground answers in the live cluster when one is reachable. Field names, defaults, and available APIs vary by version. Prefer kubectl explain . --recursive, kubectl api-resources, and kubectl version over memory — they are generated from the running server's OpenAPI and cannot be stale. If no cluster is reachable, say which Kubernetes version your answer assumes.
2. Read-only first. Diagnose with get, describe, logs, events before proposing any mutation. When you do mutate, preview first: kubectl diff -f or --dry-run=server (server-side dry-run runs real admission/validation).
3. Respect the management style of each object. Mixing kubectl apply with edit/set/patch on the same object silently loses fields on the next apply (three-way merge against last-applied). If the cluster is GitOps-managed (Flux/Argo CD field managers visible in managedFields), hand-edits get reverted at the next reconcile — direct the change to the source repo instead.
4. Never invent flags or fields. If unsure, check kubectl --help or kubectl explain. Cite the exact command you verified.
5. Be cautious with destructive operations. delete --grace-period=0 --force, drain, taints with NoExecute, and namespace deletion all have blast radius — state the consequence before giving the command, and prefer the gentler alternative (e.g. rollout restart over deleting pods, cordon before drain).

Debugging playbooks

Work top-down: kubectl get (what's wrong) → kubectl describe (events tell you why) → kubectl logs (app's view) → deeper tools. Most answers are in describe events.

Pod Pending — it hasn't been scheduled. kubectl describe pod events show the filter that failed: insufficient CPU/memory (check kubectl top nodes and pod requests), unsatisfiable nodeSelector/affinity, untolerated taints (kubectl describe node | grep -A3 Taints), or an unbound PVC (kubectl get pvc — Pending PVC = StorageClass/provisioner issue, or WaitForFirstConsumer deadlock).

CrashLoopBackOff — container starts then dies; backoff doubles 10s→5min. Sequence:

1. kubectl logs --previous — the crashed container's output (current logs are the new attempt, often empty).
2. kubectl describe pod → Last State / exit code: 137 = OOMKilled (raise memory limit) or SIGKILL after grace period; 1/2 = app error; 126/127 = bad command/missing binary.
3. Check whether a liveness probe is killing a healthy-but-slow app (events show "Liveness probe failed"). Fix = startup probe or bigger initialDelaySeconds, not more retries.
4. Config errors: missing ConfigMap/Secret keys, bad env, wrong args — describe shows CreateContainerConfigError.

ImagePullBackOff / ErrImagePull — describe events contain the registry error verbatim: typo'd image/tag, missing imagePullSecrets (private registry: needs a kubernetes.io/dockerconfigjson secret referenced in the pod spec), or wrong architecture.

Service not reachable — almost always selector/port mismatch or no ready endpoints:

1. kubectl get endpointslices -l kubernetes.io/service-name= — empty? The Service selector matches no Ready pods. Compare spec.selector to pod labels; check pod readiness (a Ready=False pod is removed from endpoints by design).
2. Port chain: client → Service port → targetPort → containerPort. targetPort must match what the app actually listens on (test with kubectl exec -- wget -qO- localhost:).
3. DNS: kubectl run -it --rm dbg --image=busybox:1.36 --restart=Never -- nslookup ..svc.cluster.local.
4. NetworkPolicy: any policy selecting the server pods makes traffic deny-by-default for the directions it lists — check kubectl get netpol -n .

Stuck Terminating — a finalizer isn't being cleared: kubectl get -o jsonpath='{.metadata.finalizers}'. Fix the controller responsible; patching the finalizer away is last resort (orphans external resources).

OOMKilled / throttling — memory over limit = kill; CPU over limit = throttle (slow, not dead). kubectl top pod --containers vs limits. QoS matters under node pressure: BestEffort (no requests) is evicted first — always set requests.

Node NotReady — kubectl describe node: pressure conditions (Memory/Disk/PID), kubelet heartbeats (Lease objects), then node-level inspection via kubectl debug node/ -it --image=busybox (host filesystem at /host).

Writing manifests — checklist

Apply this to every manifest you produce or review; each item prevents a real production failure mode:

• Current apiVersion (apps/v1, batch/v1, networking.k8s.io/v1) — verify against kubectl api-resources if a cluster is available; beta APIs get removed on upgrades.
• Pinned image tags — :latest makes rollbacks and node cache behavior nondeterministic.
• Resources: always requests (scheduling + QoS); set memory limits (OOM protection); CPU limits optional (throttling tradeoff).
• Probes: readiness (gates traffic) almost always; liveness only if a restart actually fixes the failure mode, and never probing downstream dependencies; startup probe for slow boots.
• Labels: consistent app.kubernetes.io/name|instance|component used by both selector and Service; selectors are immutable on Deployments — choose once.
• Security: runAsNonRoot: true, allowPrivilegeEscalation: false, readOnlyRootFilesystem where possible, drop capabilities; needed for restricted Pod Security level (full pattern in references/security.md).
• Dedicated ServiceAccount per app — never the namespace default SA (RBAC granted to it leaks to every pod); automountServiceAccountToken: false if the app doesn't use the Kubernetes API.
• Secrets stay in Secrets — never credentials in ConfigMaps, plain env values, or images.
• Graceful shutdown: app handles SIGTERM, or preStop sleep for connection draining; terminationGracePeriodSeconds sized to real shutdown time.
• Right workload kind: stateless → Deployment; stable identity/storage → StatefulSet; per-node → DaemonSet; run-to-completion → Job/CronJob.

Reference files — read when the task goes deeper

| File | Read when the task involves | |---|---| | references/kubectl.md | kubectl internals: apply's three-way merge, server-side apply flags, kubectl debug modes, JSONPath/custom-columns scripting, output formats, plugins, scripting conventions | | references/workloads-scheduling.md | Pod lifecycle details (phases, conditions, probe tuning, termination), workload controller behavior (Deployment rollouts, StatefulSets, Jobs), scheduling (affinity, taints, topology spread, priority/preemption, PDBs, eviction) | | references/networking-storage.md | Service types and DNS, EndpointSlices, Ingress vs Gateway API, NetworkPolicy semantics, PV/PVC lifecycle, StorageClasses, access modes, CSI, ConfigMaps/Secrets consumption details | | references/api-machinery.md | API groups/versioning/deprecation, ObjectMeta semantics (resourceVersion, generation, finalizers, ownerReferences), watches, server-side apply field ownership, the apiserver request path (authn → RBAC → admission), RBAC objects and rules, authoring CRDs/operators | | references/helm.md | Helm: chart anatomy, templating, values precedence, hooks, upgrade/rollback/stuck-release recovery, OCI registries, render debugging (helm template/diff), Helm under GitOps | | references/security.md | Hardening: Pod Security Standards and the securityContext that passes restricted, dedicated ServiceAccounts, least-privilege RBAC YAML + audit commands, NetworkPolicy patterns (default-deny baseline), secrets hygiene (SOPS/sealed-secrets/ESO), image security, ResourceQuota/LimitRange, manifest security checklist | | references/gitops.md | GitOps: Flux and Argo CD operations and diagnosis, repo layout, kustomize (bases/overlays, patches, configMapGenerator hash rollouts, images transformer), secrets-in-git options, progressive delivery, drift detection | | references/versioning-and-sources.md | Official source baseline, refresh checklist, and version-sensitive answering rules |

Scripts

All scripts are read-only and safe to run against a live cluster; each takes -h/--help. Use them to gather grounded evidence before answering — they automate the playbooks above.

• scripts/k8s-context-check.sh — local tool versions, current Kubernetes context, server/API discovery, and key kubectl explain probes before version-sensitive troubleshooting.
• scripts/k8s-diagnose.sh — triages unhealthy pods top-down (get → events → previous logs), classifying Pending, CrashLoopBackOff, ImagePullBackOff, CreateContainerConfigError, OOMKilled, stuck Terminating, and Unschedulable, with the recommended next step for each. Run this first on any "my pod is broken" question.
• scripts/k8s-manifest-lint.sh — offline static review of manifest YAML (file or stdin, e.g. helm template ... | k8s-manifest-lint.sh) against the Writing manifests checklist: deprecated apiVersions, :latest/untagged images, missing requests/limits/probes, weak securityContext, default ServiceAccount, plaintext secrets in env.
• scripts/k8s-rbac-check.sh — RBAC audit. With a ServiceAccount it reports effective permissions (auth can-i --list), high-risk verb checks, and the bindings that grant them — the fast path for Forbidden errors and over-permissioned workloads. With no argument it scans the cluster for risky bindings (cluster-admin/admin/edit).
• scripts/k8s-net-debug.sh — Service connectivity triage: port→targetPort chain, Ready endpoint count, selector-vs-pod-label matching, NetworkPolicies selecting the backend pods, and the DNS FQDN to test. Run this for Service not reachable.

Read the relevant file before answering in-depth questions in its area — they contain field-level specifics (exact defaults, version notes, failure modes) that make the difference between a plausible answer and a correct one.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: glapsfun
• Source: glapsfun/cnative-skills
• License: MIT

Install and usage instructions live in the source repository linked above.