CKA: Workloads & Scheduling

Deploy, Scale, Roll Back

Fast, declarative workflows for the exam:

kubectl create deploy web --image=nginx:1.25 --dry-run=client -o yaml > d.yaml
kubectl scale deploy web --replicas=4
kubectl set image deploy/web web=nginx:1.26
kubectl rollout status deploy/web
kubectl rollout history deploy/web
kubectl rollout undo deploy/web            # or --to-revision=N

Multi-container patterns: sidecars (logging, proxy), init containers (setup before the main container starts). CronJobs schedule batch work (schedule, concurrencyPolicy); StatefulSets give stable identity and ordered rollout for stateful services; DaemonSets ensure one pod per node — know which controller fits the requirement before writing YAML.

Scheduling Controls

Pin, spread, and tolerate deliberately:

spec:
  nodeSelector: { disktype: ssd }
  tolerations:
  - { key: spot, operator: Exists, effect: NoSchedule }
  topologySpreadConstraints:
  - maxSkew: 1
    topologyKey: topology.kubernetes.io/zone
    whenUnsatisfiable: ScheduleAnyway
    labelSelector:
      matchLabels: { app: web }

Taints repel (set with kubectl taint nodes n1 spot=true:NoSchedule); tolerations permit specific pods through; labels route with selectors (kubectl label node n1 disktype=ssd). Protect voluntary disruption:

minAvailable: 3   # or maxUnavailable: 1  → PDB

so kubectl drain waits instead of emptying the service. Understand Pending: read kubectl describe pod — taints, insufficient CPU/memory, affinity, or PVC binding are the usual causes.

Triage Order for Broken Workloads

Run the same sequence every time:

1. Pending → kubectl describe pod events: unschedulable (taints/resources/affinity), volume problems
2. ImagePullBackOff → image name/tag, registry auth, imagePullSecrets
3. CrashLoopBackOff → kubectl logs <pod> --previous, then describe for exit codes (137 = OOM, 1 = app error)
4. Running but not Ready → readiness probe path/port and app dependencies; check kubectl get endpoints
5. Stuck Terminating → finalizers, hung preStop hooks, grace period (--force --grace-period=0 only as a last resort)

Also: kubectl get events --sort-by=.lastTimestamp for the freshest signals, and kubectl top pod once metrics exist. Interview framing: narrate which signal you read at each step and the specific fix it dictates — interviewers score the diagnostic order as heavily as the answer.