On 2022-Jan-12, Jaime Casanova wrote:
> I found two crashes, actually I found them on the original patch Álvaro
> sent on november but just checked that those already exists.
>
> I configured with:
>
> CFLAGS="-ggdb -Og -g3 -fno-omit-frame-pointer" ./configure --prefix=/opt/var/pgdg/15/merge --enable-debug
--enable-depend--enable-cassert --with-llvm --enable-tap-tests --with-pgport=54315
>
> And tested on the regression database.
>
> Attached the SQL files for the crashes and its respective stacktraces.
> FWIW, the second crash doesn't appear to be caused by the MERGE patch
> but I cannot trigger it other way.
Thanks for this! The problem in the first crash was that when
partitioned tables are being used and the topmost one has a tuple
descriptor different from the partitions, we were doing the projection
to the partition's slot using the root's tupledesc and a targetlist
written for the root. The reason this crashed in such ugly way is that
in this case the parent has 3 columns (2 dropped) while the partitions
only have one, so the projection was trying to write to an attribute
that didn't exist.
I fixed it by making all NOT MATCHED actions use the root table's
descriptor and slot.
This change fixes both your reported crashes. I didn't look closely to
see if the second one is caused by exactly the same issue.
--
Álvaro Herrera Valdivia, Chile — https://www.EnterpriseDB.com/
Tom: There seems to be something broken here.
Teodor: I'm in sackcloth and ashes... Fixed.
http://archives.postgresql.org/message-id/482D1632.8010507@sigaev.ru