Connection.close(timeout=) waits forever on a pending cancel when the server never acknowledges it

[developper89]

Summary

When a statement times out (command_timeout) while the server, or a pooler in front of it, is frozen, asyncpg requests a cancel and then every later operation on that connection, including close(timeout=...), awaits the cancel acknowledgement with no bound. The timeout argument of close() does not cover that wait, and a subsequent transport loss does not resolve it either, so the connection can never be closed gracefully and any caller that awaits close() hangs indefinitely.

Versions

• asyncpg 0.30.0 and 0.31.0 (same code shape in both)
• Python 3.12.3, Linux
• Observed through SQLAlchemy 2.0.52's asyncpg dialect, which calls Connection.close(timeout=2) when invalidating a connection after a TimeoutError, but the behaviour is asyncpg's.

Where in the source (0.31.0)

• asyncpg/protocol/protocol.pyx, close(self, timeout): awaits self.cancel_sent_waiter and then if self.cancel_waiter is not None: await self.cancel_waiter before the part that is guarded by timeout.
• _request_cancel() (called from _on_timeout()) creates cancel_waiter; it is resolved only by a ReadyForQuery arriving on the original socket.
• _handle_waiter_on_connection_lost() and _on_connection_lost() resolve self.waiter only; cancel_waiter is left pending when the transport is lost.
• abort() returns early when self.closing is already set, so cancelling a stuck close() from outside and then calling Connection._abort() does not close the transport.

Reproduction

1. Run PostgreSQL behind pgbouncer (transaction pooling), or plain PostgreSQL.
2. Open a connection with command_timeout=5, run SELECT pg_sleep(40).
3. While it runs, freeze the server process (podman pause / kill -STOP on postgres, or on pgbouncer).
4. The statement raises asyncio.TimeoutError after 5 s and asyncpg starts a cancel task.
5. Now await conn.close(timeout=2): it never returns while the freeze lasts. If the frozen side is later closed by a pooler timeout (pgbouncer query_timeout closes the client socket), close() still never returns because the transport loss resolves only the query waiter.

Observed with an asyncio task-stack watchdog: the caller sits in Connection.close → protocol.close → await self.cancel_waiter, and the _cancel task sits in connect_utils awaiting the cancel connection's on_disconnect, for as long as the server stays frozen (minutes; unbounded).

Expected

• close(timeout=t) should bound the wait for the cancel acknowledgement by t (or by the connection's command_timeout) and fall back to aborting the transport.
• _on_connection_lost() should resolve cancel_waiter (with the same connection-lost exception it uses for waiter) so that a lost transport cannot leave a permanently pending cancel.

Workaround we use

An application-level guard that, on TimeoutError/CancelledError, runs asyncio.wait_for(conn.close(timeout=g), g) and on expiry calls conn.terminate() followed by an explicit conn._transport.abort(), because terminate() alone leaves the socket open once close() has marked the protocol as closing.