On Tue, Mar 04, 2025 at 04:53:14PM -0800, Jacob Champion wrote:
> I'll work on a fix, but it probably won't be fast since I need to
> learn more about the injection points architecture. The test may need
> to be disabled, or the patch backed out, depending on how painful the
> flake is for everybody.
Oops, missed these failures.. So we have a race condition where we
are trying to wake up a point that's not waiting yet, because there is
a small window between the moment when the backend entry is marked as
"starting" and the injection point wait.
What this is telling us is that we should change the query scanning
pg_stat_activity for a PID of a backend in 'starting' state so as we
also check the wait_event init-pre-auth, as this is reported when
using injection point waits. The attached should be enough to take
care of this race condition.
--
Michael