One approach that sidesteps the whole problem: design for fully synchronous, stateless requests from the start so there's nothing to queue.

I did this for a financial calculator API — every request is pure computation, inputs in, result out, nothing persisted. No Redis, no workers, no task table, no locking. The response is ready before a user would notice a queue anyway (sub-50ms).

Obviously only works when tasks complete in milliseconds. But figassis's pattern of "starts simple, then incrementally grows into a small job system anyway" often happens because the initial scope could have been fully synchronous — the async complexity creeps in before it's actually needed.

Worth asking first: does this task genuinely have to be async, or is it just easier to model it that way?