Limitations

Jobs can execute more than once

A crashed or timed-out worker can leave a claim that another worker later retries. If the original worker is still executing, both can perform the same operation. A claim also cannot make database updates and an external API call one atomic transaction.

Choose task timeouts above the maximum expected duration, within defaultRequeueTimeout, and make tasks safe to retry. Use stable operation IDs, database constraints, or the target service's idempotency support where appropriate.

Task concurrency

A task's unique flag restricts concurrent execution of that task type. rate applies within each worker, and costs limits the per-server cost budget. These controls do not guarantee exactly-once effects and do not provide arbitrary per-tenant concurrency limits.

The unique option passed to createJob() is separate: it deduplicates pending jobs with the same task and reference. It does not prevent a completed operation from being queued again.