| name | kubernetes |
| description | Kubernetes workload discipline — probes, resources, RBAC, disruption budgets. CNCF standard, not vendor; cluster provisioning belongs to infra. |
Skill: Kubernetes
K8s is a CNCF open standard — ubiquitous enough to be discipline, not vendor. Boundary: workload operation lives here (dev-ops); cluster lifecycle/provisioning is infra (platform-architecture, infra-patterns). Managed-flavor specifics from memory-long §stack.
Workload non-negotiables
- Requests and limits on every container — no requests = scheduler flying blind = noisy-neighbor outages; memory limit ≈ request (OOMKill beats node eviction), CPU limit optional-by-policy (throttling trade-off documented).
- Liveness ≠ readiness: liveness = "restart me" (deadlock detection ONLY — a liveness probe hitting the DB restarts your app when the DB blips) · readiness = "don't route to me yet" (dependencies belong here) · startup probes for slow boots instead of generous liveness delays.
- Deployments declarative via GitOps (see
gitops) — kubectl edit/apply by hand in prod = drift incident · rollout strategy explicit (maxSurge/maxUnavailable are release-engineering choices).
- PodDisruptionBudgets on anything that matters — node drains and upgrades WILL happen; a PDB is how your availability survives ops (see
resilience).
Isolation & access
Namespaces are the unit of tenancy: quotas + RBAC + network policies per namespace · RBAC least-privilege per service account — no default SA with permissions, no cluster-admin for workloads, ever (see infra-security identity rules) · NetworkPolicies deny-by-default east-west (see network-design) · secrets as mounted refs from a real secret store, not bare base64 manifests in git.
State & scaling
Stateless by default; StatefulSets only for genuinely stateful workloads (and prefer managed data services — running databases on k8s is an ADR, not a habit) · HPA on metrics that mean something (queue depth > CPU for workers) · anti-affinity for replicas of the same service — three replicas on one node is one replica with extra steps.
Verification
kubectl diff-style dry-run in CI before any manifest merges · new workload checklist: probes, resources, PDB, SA, network policy, non-root (see containers) · post-deploy: rollout status + RED metrics green before the report closes (see observability).
