A restart policy tells Kubernetes whether to replace a pod’s container after its process exits. Set `restartPolicy` in the pod spec: `Always` restarts after any exit and is the default; `OnFailure` restarts only after a non-zero exit code; `Never` does not restart. The pod-level setting applies to all its regular containers, not to each one separately. section 6.2.1
Let a Job retry a failed task
In a Job manifest, put `restartPolicy: OnFailure` under `spec.template.spec`, alongside the container definition. A Job pod must explicitly use `OnFailure` or `Never`: the default `Always` would prevent a successfully finished pod from completing. With `OnFailure`, the Kubelet replaces a failed container in the same pod; once the task succeeds, the pod can complete. section 18.1.1 section 18.1.3
Listing 18.1 A Job manifest for running a single task
apiVersion: batch/v1 #1
kind: Job #1
metadata:
name: quiz-init
labels:
app: quiz
task: init
spec:
template: #2
metadata: #3
labels: #3
app: quiz #3
task: init #3
spec:
restartPolicy: OnFailure #4
initContainers: #5
- name: init #5
image: mongo:5 #5
command: #5
- sh #5
- -c #5
- | #5
mongosh mongodb://quiz-0.quiz-pods.kiada.svc.cluster.local \ #5
--quiet --file /dev/stdin <<EOF #5
#5
# MongoDB code that initializes the replica set #5
# Refer to the job.quiz-init.yaml file to see the actual code #5
#5
EOF #5
containers: #6
- name: import #6
image: mongo:5 #6
command: #6
- mongoimport #6
- mongodb+srv://quiz-pods.kiada.svc.cluster.local/kiada?tls=false #6
- --collection #6
- questions #6
- --file #6
- /questions.json #6
- --drop #6
volumeMounts: #6
- name: quiz-data #6
mountPath: /questions.json #6
subPath: questions.json #6
readOnly: true #6
volumes:
- name: quiz-data
configMap:
name: quiz-data
The Job manifest shows exactly where `restartPolicy: OnFailure` belongs in a pod template. section 18.1.1
“Restart” does not revive the old container: Kubernetes discards it and creates another. Data written only to its filesystem is lost. Repeated failures trigger increasing delays between attempts, which can reach five minutes; `CrashLoopBackOff` can appear while the container waits. If a regular container restarts, the pod’s init containers normally do not run again. section 6.2.1
The choice also determines who handles a failed Job task. With `OnFailure`, the Kubelet retries the container inside its existing pod. With `Never`, that pod fails and the Job controller creates another pod, possibly on a different node. A failing init container is likewise retried under `OnFailure` or `Always`; under `Never`, later containers never start. section 18.1.3 section 6.4.1
Do not confuse the pod’s policy with the `restartPolicy: Always` that can appear on an individual init container. In Kubernetes 1.28 and later, that setting marks a native sidecar, which stays running alongside the pod’s regular containers. section 5.5.4
