Closed matthewwardrop closed 2 months ago
Sadly, I am attempting to figure this too. I was reproducing it earlier today, but my simple repo test case, just failed to reproduce it sigh...
Ugh. Heisenbugs are the worst.
Of course, if it continues to not reproduce more globally, I'd take the win ;).
@matthewwardrop Are you apt
installing anything onto your GHA runner? Hopefully I can isolate a simple repo shortly
Hi @akrherz ! I experience these issues when locally installing into an isolated conda environment. We do install some pip dependencies as well, but these are for proprietary purposes so I stripped them out of the conda list. The pip deps didn't override the psycopg package.
sweet, I reproduced it in a simple repo, whew. But, I can not reproduce locally on linux 64, sigh.
I have also now reproduced this on Github Actions with psycopg2 2.9.9, so a rebuild earlier today of it did not help :(
related: conda-forge/openssl-feedstock#144
kindly wondering if @h-vetinari and/or @hmaarrfk have thoughts on this issue?
You can see from the environment in the OP that it doesn't have OpenSSL 3.2 in it:
openssl 3.1.4 hd590300_0 conda-forge
So this is very likely an unrelated issue. Also, to pick up the fix for https://github.com/conda-forge/postgresql-feedstock/issues/178, you need at least build number 1 or higher, whereas the environment in the OP has:
libpq 16.1 hfc447b1_0 conda-forge
^
too old
My psql reproducer here has
postgresql 16.1 h7387d8b_4 conda-forge
It's not completely impossible that the patch backported in https://github.com/conda-forge/postgresql-feedstock/pull/181 (by touching the config) created a kind of soft ABI-break. Something compiled against unpatched postgresql might pick up a different config than the new one, with correspondingly different behaviour.
This is just hypothesis for now, but if so, we should rebuild this feedstock (i.e. compile against newest libqp) and see if the error persists. If so, we need a repodata patch to ensure that previous builds don't use the patched libpq (e.g. by simply adding openssl <3.2
), and do the same for the v2 branch.
CC @conda-forge/postgresql, resp. @ocefpaf for merging https://github.com/conda-forge/postgresql-feedstock/pull/181
PS. Written before @akrherz's last comment but less important now:
It's not completely impossible that the patch [...] created a kind of soft ABI-break
This is looking more and more likely IMO.
Oh... My apologies. I may also have reverted the update to OpenSSL 3.2 and verified that it solved things, but forgot to roll the packages forward again. 3.1.* works fine here. It's updating to 3.2 that causes the problem.
On Mon, Dec 4, 2023, 21:43 h-vetinari @.***> wrote:
It's not completely impossible that the patch [...] created a kind of soft ABI-break
This is looking more and more likely https://github.com/conda-forge/postgresql-feedstock/issues/191#issuecomment-1840051154 IMO.
— Reply to this email directly, view it on GitHub https://github.com/conda-forge/psycopg2-feedstock/issues/94#issuecomment-1840052389, or unsubscribe https://github.com/notifications/unsubscribe-auth/AAA6P3RCILN6U4HAI7BLCNLYH2YB7AVCNFSM6AAAAABAGZKP7OVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTQNBQGA2TEMZYHE . You are receiving this because you were mentioned.Message ID: @.***>
build 6 of postgresql/libpq 16.1 should fix this issue. I am no longer able to reproduce it.
@matthewwardrop Are you still having this issue?
I'm closing this issue as it seems like it might be fixed. Please re-open if not.
Solution to issue cannot be found in the documentation.
Issue
Upstream
postgres
had a bug in its usage of openssl that caused it segfault with OpenSSL 3.2.0. This was fixed: https://github.com/conda-forge/postgresql-feedstock/issues/178 . However, I still see the segfault when usingpsycopg2
. Is the binary statically linked or some such?Installed packages
Environment info