One Pod per node, a task that finishes, or a task on a schedule.

DaemonSet

n1n2n3n4

Exactly one Pod on every (or every matching) node; a new node gets one automatically.

Log shippers, node-exporter, CNI and CSI agents.

Job

01234

Runs Pods until completions succeed. parallelism sets how many at once; backoffLimit caps retries with exponential backoff.

completionMode: Indexed gives each Pod an index for sharded batch work.

CronJob

02:0002:00+1d

Creates a Job on a cron schedule. Always set concurrencyPolicy.

Allow (default) may overlap, Forbid skips, Replace cancels the old run.

The classic mistake

A one-time migration run as a Deployment is restarted forever: restartPolicy: Always treats a clean exit as a crash. Use a Job (OnFailure or Never).

Backups must never overlap

A nightly backup that occasionally runs long can overlap with the next run under Allow. Use Forbid and set startingDeadlineSeconds.

DaemonSets run everywhere; Jobs run to completion

A DaemonSet ensures one Pod runs on every node, or every node matching a selector, and automatically adds one when a node joins. It is the right object for node-level infrastructure: log collectors, metrics exporters, CNI plugins, CSI node drivers, security agents. It is the wrong object for application replicas.

A Job runs Pods to completion. completions says how many successful Pods are needed, parallelism how many may run at once, backoffLimit how many failures are tolerated with exponential backoff, and activeDeadlineSeconds a total time cap. Indexed completion mode gives each Pod a unique index through JOB_COMPLETION_INDEX, ideal for splitting a dataset into shards without an external work queue.

A CronJob creates Jobs on a schedule. Its most important field is concurrencyPolicy. The default Allow lets a slow run overlap with the next, which can cause two backups hammering one database. Forbid skips the new run while one is active; Replace cancels the old one. Also consider startingDeadlineSeconds to bound how late a missed run may start, and the history limits to avoid piling up finished Jobs.

The classic misuse is running a one-time task as a Deployment. A Deployment's Pods always use restartPolicy Always, so a successfully exiting container is restarted forever. Jobs understand that exit code zero means done.