mosquito / aiohttp-tiny-mcp / 35507528139

94%

Build Type

push

github

Committed by mosquito

Commit Message

Cancel a relay task again when its first cancellation was lost

Closing a 2024-11-05 stream hung now and then on Python 3.10 and 3.11
with the Postgres backend: psycopg_pool waits for a connection with
asyncio.wait_for, which before 3.12 drops a cancellation that arrives in
the loop turn the wait completes. The second relay task, cancelled
first, hands its connection back at exactly that moment, so the first
task's cancel is swallowed and the join never returns; server shutdown
then waits for the handler past the test timeout.

tasks.stop cancels, waits a grace period, and cancels again whatever is
still running, until every task has ended. Every cancel-then-await site
uses it: both relay joins, run_until_disconnect, and stdio shutdown. A
hub that swallows the first cancellation reproduces the hang for the
HTTP+SSE stream and the GET stream.

Coverage Stats

26 of 26 new or added lines in 4 files covered. (100.0%)

45 existing lines in 7 files now uncovered.

5240 of 5561 relevant lines covered (94.23%)

0.94 hits per line