[efeesrefinal]

Fixes #650

Since 0.17.0 completeJob and failJob are fire-and-forget, so a worker that finishes a job in a named queue asks for its next job before the completion has been committed. The queue is still locked, getJob returns nothing and the worker sleeps for pollInterval. Only jobs:insert nudges idle workers, so once inserts stop the queue advances once per poll.

This change nudges the pool after completing or failing jobs that belong to a named queue has been committed, once per released queue, in the same way a jobs:insert notification does. It covers the batched and unbatched completion paths, and with localQueue the nudge becomes a pulse(). It does nothing while the pool is shutting down. Jobs without a queue are unaffected.

The nudge is local to the pool that completed the job. Other worker processes still wait for their poll or a notification, the same as before 0.17.0. I kept it local because sending a pg_notify per completion seemed too heavy for what is a wake-up hint, but I'm happy to change that if you'd prefer.

Tests: added a test.each to main.runTaskList.test.ts that inserts 5 jobs into one named queue in a single statement and runs a pool with pollInterval: 60_000, with and without localQueue. Both cases time out on main (sleepUntil 5 s) and finish in about 200 ms with this change. The full suite passes (23 suites, 107 tests) on Node 22 and PostgreSQL 17, and eslint and prettier are clean on the changed files. I added a patch changeset.

[benjie]

Thanks for the PR! This follows pretty much the pattern I was intending to implement for this: count the number of jobs that were released that had queues, and then nudge that many workers (via the pool). I think we should move a few parts of the logic to more central locations, but other than that this looks good 👍 (I've not yet reviewed the test/changeset)

[efeesrefinal]

Thanks for the review! Both done in d875b75:

• The _active / _shuttingDown check now lives in WorkerPool.nudge, so the helper only counts the released queues, and the try/catch is gone.
• worker.nudge() no longer asserts: if the worker is no longer active it logs at debug level and returns false, so the pool moves on to the next worker.

The full suite still passes, and the two new tests still fail on main.

[benjie]

I went with a slightly different approach that eliminates additional loops and promises and reduces the amount of code needed; I also trimmed down some of the verbosity of the LLM. WDYT?

[efeesrefinal]

Thanks, that's much cleaner, and having the nudge right after the query that releases the lock makes more sense. Sorry for leaving the unused import behind, and thanks for fixing it.