Обсуждение: Sporadic connection-setup-related test failures on Cygwin in v15-

Hello hackers,

A recent lorikeet (a Cygwin animal) failure [1] revealed one more
long-standing (see also [2], [3], [4]) issue related to Cygwin:
  SELECT dblink_connect('dtest1', connection_parameters());
- dblink_connect
-----------------
- OK
-(1 row)
-
+ERROR:  could not establish connection
+DETAIL:  could not connect to server: Connection refused

where inst/logfile contains:
2024-07-16 05:38:21.492 EDT [66963f67.7823:4] LOG:  could not accept new connection: Software caused connection abort
2024-07-16 05:38:21.492 EDT [66963f8c.79e5:170] pg_regress/dblink ERROR:  could not establish connection
2024-07-16 05:38:21.492 EDT [66963f8c.79e5:171] pg_regress/dblink DETAIL:  could not connect to server: Connection
refused
         Is the server running locally and accepting
         connections on Unix domain socket "/home/andrew/bf/root/tmp/buildfarm-DK1yh4/.s.PGSQL.5838"?

I made a standalone reproducing script (assuming the dblink extension
installed):
numclients=50
for ((i=1;i<=1000;i++)); do
echo "iteration $i"

for ((c=1;c<=numclients;c++)); do
cat << 'EOF' | /usr/local/pgsql/bin/psql >/dev/null 2>&1 &

SELECT 'dbname='|| current_database()||' port='||current_setting('port')
   AS connstr
\gset

SELECT * FROM dblink('service=no_service', 'SELECT 1') AS t(i int);

SELECT * FROM
dblink(:'connstr', 'SELECT 1') AS t1(i int),
dblink(:'connstr', 'SELECT 2') AS t2(i int),
dblink(:'connstr', 'SELECT 3') AS t3(i int),
dblink(:'connstr', 'SELECT 4') AS t4(i int),
dblink(:'connstr', 'SELECT 5') AS t5(i int);
EOF
done
wait

grep -A1 "Software caused connection abort" server.log && break;
done

which fails for me as below:
iteration 318
2024-07-24 04:19:46.511 PDT [29062:6][postmaster][:0] LOG:  could not accept new connection: Software caused connection

abort
2024-07-24 04:19:46.512 PDT [25312:8][client backend][36/1996:0] ERROR:  could not establish connection

The important fact here is that this failure is not reproduced after
7389aad63 (in v16), so it seems that it's somehow related to signal
processing. Given that, I'm inclined to stop here, without digging deeper,
at least until there are plans to backport that fix or something...

[1] https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=lorikeet&dt=2024-07-16%2009%3A18%3A31 (REL_13_STABLE)
[2] https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=lorikeet&dt=2022-07-21%2000%3A36%3A44 (REL_14_STABLE)
[3] https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=lorikeet&dt=2023-07-06%2009%3A19%3A36 (REL_12_STABLE)
[4] https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=lorikeet&dt=2022-02-12%2001%3A40%3A56 (REL_13_STABLE, 
postgres_fdw)

Best regards,
Alexander

On Thu, Jul 25, 2024 at 1:00 AM Alexander Lakhin <exclusion@gmail.com> wrote:
> The important fact here is that this failure is not reproduced after
> 7389aad63 (in v16), so it seems that it's somehow related to signal
> processing. Given that, I'm inclined to stop here, without digging deeper,
> at least until there are plans to backport that fix or something...

+1.  I'm not planning to back-patch that work.  Perhaps lorikeet
could stop testing releases < 16?  They don't work and it's not our
bug[1].  We decided not to drop Cygwin support[2], but I don't think
we're learning anything from investigating that noise in the
known-broken branches.

[1] https://sourceware.org/legacy-ml/cygwin/2017-08/msg00048.html
[2] https://www.postgresql.org/message-id/5e6797e9-bc26-ced7-6c9c-59bca415598b%40dunslane.net

24.07.2024 23:58, Thomas Munro wrote:
> +1.  I'm not planning to back-patch that work.  Perhaps lorikeet
> could stop testing releases < 16?  They don't work and it's not our
> bug[1].  We decided not to drop Cygwin support[2], but I don't think
> we're learning anything from investigating that noise in the
> known-broken branches.

Yeah, it looks like lorikeet votes +[1] for your proposal.
(I suppose it failed due to the same signal processing issue, just another
way.)

[1] https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=lorikeet&dt=2024-07-24%2008%3A54%3A07

Best regards,
Alexander

On Thu, Jul 25, 2024 at 1:00 AM Alexander Lakhin <exclusion@gmail.com> wrote:

The important fact here is that this failure is not reproduced after
7389aad63 (in v16), so it seems that it's somehow related to signal
processing. Given that, I'm inclined to stop here, without digging deeper,
at least until there are plans to backport that fix or something...

+1.  I'm not planning to back-patch that work.  Perhaps lorikeet
could stop testing releases < 16?  They don't work and it's not our
bug[1].  We decided not to drop Cygwin support[2], but I don't think
we're learning anything from investigating that noise in the
known-broken branches.

Sure, it can. I've made that change.

cheers

andrew

--
Andrew Dunstan
EDB: https://www.enterprisedb.com

25.07.2024 19:25, Andrew Dunstan wrote:
>> +1.  I'm not planning to back-patch that work.  Perhaps lorikeet
>> could stop testing releases < 16?  They don't work and it's not our
>> bug[1].  We decided not to drop Cygwin support[2], but I don't think
>> we're learning anything from investigating that noise in the
>> known-broken branches.
>
>
> Sure, it can. I've made that change.
>

Thank you, Andrew!

I've moved those issues to the "Fixed" category.

Best regards,
Alexander