Course contentsShow
← Cloud and Infrastructure
Lesson 1193 of 2,98923. Kubernetes WorkloadsFree lesson

Job Restart and Backoff Policies

Configure pod restart policies (Never, OnFailure) and backoff limits to handle failures and control retry behavior in Jobs.

Job Restart and Backoff Policies

What you'll learn: How to configure when and how Kubernetes retries failed Job pods using restart policies and backoff limits.

Understanding Job Failures

When a Job's pod fails, Kubernetes needs to know: Should it restart immediately? Wait? Give up? This is where restart policies and backoff limits come in—they're your Job's failure-handling rulebook.

Think of it like a vending machine: if your snack gets stuck, does the machine try again immediately, wait a bit between attempts, or eventually stop trying? That's exactly what these policies control.

Pod Restart Policies for Jobs

Jobs support two restart policies in the pod template:

  • Never: If the pod fails, Kubernetes creates a new pod instead of restarting the same container. The failed pod remains for debugging.
  • OnFailure: Kubernetes restarts the failed container within the same pod. No new pod is created.
apiVersion: batch/v1
kind: Job
metadata:
  name: data-processor
spec:
  template:
    spec:
      restartPolicy: OnFailure  # or Never
      containers:
      - name: processor
        image: myapp:1.0

Backoff Limits

The backoffLimit controls how many times Kubernetes will retry before marking the Job as failed. Default is 6.

spec:
  backoffLimit: 3  # Stop after 3 failed attempts

Kubernetes uses exponential backoff (10s, 20s, 40s...) between retries, capping at 6 minutes. This prevents overwhelming your system with rapid retries.

When to Use Which

  • Use Never when you need to inspect failed pods or when failures require fresh state
  • Use OnFailure for transient errors where a simple restart might work
  • Lower backoffLimit for expensive operations; increase for flaky workloads

Key Takeaway: Restart policies determine how pods retry (new pod vs container restart), while backoffLimit controls how many times Kubernetes will attempt before giving up—together they make Job failure handling predictable and controllable.