supabase / supavisor / 29034153596

77%

Build Type

push

github

Committed by web-flow

Commit Message

fix: handle client socket errors on DbHandler, sweep zombie ports, hold check-in till flush complete (#1060)

When the timeout closes ({send_timeout_close, true} (a ranch default))
are triggered in the inet backend, the underlying TCP socket is closed,
but there's no notification to the owner process, and the port stays
alive.

We must then make sure that when we get the timeout error from send, we
adequately react and the relevant processes are terminated. The
DbHandler didn't check the return value when writing to the client, and
when timing out would cause zombie ports+client handlers, which could
even hold a database session, and potentially create long transactions.
I reproduced this scenario in a test.

This doesn't apply to the upstream sockets to database, which don't use
this option.

The dead port sweeper is cherry picked from
https://github.com/supabase/supavisor/pull/1056, which was the bug
mitigation targeted at the current v2.9 release.

Also addresses a related race condition that could cause a DbHandler
that is busy flushing data to a slow client to be checked out by another
client if (and only if) the query had already completed from the
DbHandler perspective. This was fixed by making the DbHandler checkin
itself when its work is complete, instead of relying on the
ClientHandler to check it in.I don't expect this to make much of a
difference in real life, but its a more correct behaviour.

Coverage Stats

33 of 41 new or added lines in 4 files covered. (80.49%)

4 existing lines in 4 files now uncovered.

2644 of 3424 relevant lines covered (77.22%)

54383.87 hits per line