Run Batch Work with Jobs and CronJobs

Why this matters

Migrations, report generation, queue draining, nightly cleanup: production is full of work that should run, finish, and stop. Running it as a Deployment wastes a restart loop and hides the completion signal. The exam checks the distinction constantly, plus the failure knobs, because batch jobs that retry forever are how clusters melt at night.

Concepts

A Job creates pods until the completion count is met. completions says how many successful pods finish the work; parallelism says how many run at once; backoffLimit caps retries before the whole Job is marked failed. restartPolicy: Never or OnFailure is required (the default Always belongs to long-running pods and is rejected). activeDeadlineSeconds bounds total run time so a stuck batch cannot burn quota all weekend.

A CronJob wraps a Job template in a schedule. The schedule is standard cron plus an optional timezone; concurrencyPolicy decides what happens when a run overruns its slot (Allow, Forbid, Replace). successfulJobsHistoryLimit plus failedJobsHistoryLimit keep history readable instead of infinite. Suspend pauses everything without deleting the definition, which is the safe way to stop a noisy schedule during an incident.

Worked example

Nightly report, one at a time, keep three successes for inspection:

apiVersion: batch/v1
kind: CronJob
metadata:
  name: nightly-report
spec:
  schedule: "30 2 * * *"
  concurrencyPolicy: Forbid
  successfulJobsHistoryLimit: 3
  failedJobsHistoryLimit: 1
  jobTemplate:
    spec:
      backoffLimit: 2
      activeDeadlineSeconds: 600
      template:
        spec:
          restartPolicy: OnFailure
          containers:
          - name: report
            image: busybox
            command: ["sh", "-c", "echo report-$(date +%F)"]

Check with kubectl get jobs after the minute passes, then kubectl logs job/<name> for the output. Forbid plus a ten-minute deadline means a stuck run neither piles up nor burns the night.

Exam notes

restartPolicy: Always on a Job template is rejected by the API: memorize Never or OnFailure. Cron schedules run in the controller's timezone unless set, so always state yours. And history limits prune Jobs but never running pods: a Wedged run still needs deleting by hand.