Assume work can be repeated

Background jobs may be retried after a timeout or process restart. If repeating a job creates duplicate side effects, a routine recovery can become a data problem.

Use a stable operation key or a persisted state transition to recognize work that has already completed. The mechanism should match the business operation.

Test interruptions

A useful test runs the same job more than once and checks the final state, not just the response from a single execution.

Also consider a failure between an external side effect and recording success. Idempotency is most valuable at precisely that boundary.