SWE-literacy set. Enough to answer infra-ish questions in a backend interview without panicking. Not a platform-engineering reference.
Smallest deployable unit — one or more containers that share network and storage.
Pods are disposable. You rarely create a Pod directly; a Deployment / Job / DaemonSet creates them for you and replaces them when they die. IP addresses change on restart — never refer to a pod by IP.
Declarative management of replicated stateless pods.
You describe "I want N replicas of this pod template"; the controller makes it true. Rolling updates, pause/resume, and one-command rollback come free (`kubectl rollout undo`).
Stable virtual IP + DNS name in front of a changing set of pods.
Four types. ClusterIP: in-cluster only (default). NodePort: exposed on every node at a high port. LoadBalancer: cloud-provisioned external LB. ExternalName: CNAME to an external host — useful for abstracting external dependencies.
HTTP(S) routing to services based on hostname / path.
Not a resource the cluster implements on its own — requires an Ingress Controller (nginx-ingress, Traefik, cloud-native). The Ingress object is the spec; the Controller turns it into real load-balancer rules.
Non-sensitive key/value config, decoupled from the image.
Mount as env vars or as files under a volume. Changes do NOT auto-restart pods — either roll the Deployment or use a tool like reloader. Config is the right place for feature flags, environment names, tuning knobs.
Same as ConfigMap but for sensitive values.
Secrets are NOT encrypted by default — only base64-encoded. Enable encryption-at-rest in etcd, restrict RBAC to view secrets, and consider an external secrets operator for real key management.
The most-asked K8s interview topic. Each probe answers a different question — mixing them up is the classic mistake.
| Probe | Answers | On failure | Use for |
|---|---|---|---|
| Liveness | Is this container wedged? Should I kill it? | Kubelet restarts the container | For deadlock detection — process is running but making no progress. Keep it cheap; a flaky liveness probe causes restart loops |
| Readiness | Is this container ready to serve traffic? | Pod is removed from Service endpoints until it recovers | For warm-up, transient downstream outages, circuit-breaker integration. Failing a readiness probe is NOT destructive — the pod keeps running |
| Startup | Has the app finished starting? | Disables liveness / readiness until it passes; kills pod on timeout | For slow-starting apps (JVM warm-up, large caches) — lets liveness have a tight timeout in steady state without triggering during startup |
Senior insight
Senior insight
Apply a manifest change
kubectl apply -f deployment.yaml Watch the rollout progress
kubectl rollout status deployment/my-app Pause mid-rollout (hold current progress)
kubectl rollout pause deployment/my-app Resume
kubectl rollout resume deployment/my-app Rollback to previous revision
kubectl rollout undo deployment/my-app Rollback to a specific revision
kubectl rollout undo deployment/my-app --to-revision=3 See rollout history
kubectl rollout history deployment/my-app List pods in the current namespace
kubectl get pods Watch pods live
kubectl get pods -w All resources in a namespace
kubectl get all -n my-namespace Wide output (IPs, node)
kubectl get pods -o wide Full YAML of a resource
kubectl get pod my-pod -o yaml Events + config + status on a single resource
kubectl describe pod my-pod Tail logs
kubectl logs -f my-pod Logs from previous (crashed) container
kubectl logs my-pod --previous Multi-container pod — pick a container
kubectl logs my-pod -c my-container Interactive shell into a container
kubectl exec -it my-pod -- sh Port-forward to your local machine
kubectl port-forward my-pod 8080:80 Resource usage (requires metrics-server)
kubectl top pods metrics-server must be installed
Context + namespace at a glance
kubectl config current-context && kubectl config view --minify | grep namespace Out of scope for a SWE interview
These topics exist, matter, and are important — but they sit in platform-engineering territory. A software engineer is rarely asked to design them from scratch in a loop.