09.02.2024 21:59, Andres Freund wrote:
https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=kestrel&dt=2024-02-04%2001%3A53%3A44
) and saw that it's not checkpointer, but walsender is hanging:
How did you reproduce this?
As kestrel didn't produce this failure until recently, I supposed that the
cause is the same as with subscription/031_column_list — longer test
duration, so I ran this test in parallel (with 20-30 jobs) in a slowed
down VM, so that one successful test duration increased to 100-120 seconds.
And I was lucky enough to catch it within 100 iterations. But now, that we
know what's happening there, I think I could reproduce it much easily,
with some sleep(s) added, if it would be of any interest.
So it's the issue that we wait effectively forever to to send a FATAL. I've
previously proposed that we should not block sending out fatal errors, given
that allows clients to do prevent graceful restarts and a lot of other things.
Yes, I had demonstrated one of those unpleasant things previously too:
https://www.postgresql.org/message-id/91c8860a-a866-71a7-a060-3f07af531295%40gmail.com
Best regards,
Alexander
