On Mon, Jul 10, 2023 at 09:01:37PM -0500, Justin Pryzby wrote:
> An instance compiled locally, without assertions, failed like this:
>
...
>
> => REINDEX was running, with parallel workers, but deadlocked with
> ANALYZE, and then crashed.
>
> It looks like parallel workers are needed to hit this issue.
> I'm not sure if the issue is specific to extended stats - probably not.
>
> I reproduced the crash with manual REINDEX+ANALYZE, and with assertions (which
> were not hit), and on a more recent commit (1124cb2cf). The crash is hit about
> 30% of the time when running a loop around REINDEX and then also running
> ANALYZE.
>
> I hope someone has a hunch where to look; so far, I wasn't able to create a
> minimal reproducer.
I was able to reproduce this in isolation by reloading data into a test
instance, ANALYZEing the DB to populate pg_statistic_ext_data (so it's
over 3MB in size), and then REINDEXing the stats_ext index in a loop
while ANALYZEing a table with extended stats.
I still don't have a minimal reproducer, but on a hunch I found that
this fails at 5764f611e but not its parent.
commit 5764f611e10b126e09e37fdffbe884c44643a6ce
Author: Andres Freund <andres@anarazel.de>
Date: Wed Jan 18 11:41:14 2023 -0800
Use dlist/dclist instead of PROC_QUEUE / SHM_QUEUE for heavyweight locks
I tried compiling with -DILIST_DEBUG, but that shows nothing beyond
segfaulting, which seems to show that the lists themselves are fine.
--
Justin