I wrote: >> diff -w -U3 >> .../src/interfaces/ecpg/test/expected/thread-descriptor.stderr >> .../src/interfaces/ecpg/test/results/thread-descriptor.stderr >> --- .../src/interfaces/ecpg/test/expected/thread-descriptor.stderr >> 2019-12-04 16:05:46 +0300 >> +++ .../src/interfaces/ecpg/test/results/thread-descriptor.stderr >> 2020-10-18 20:20:27 +0300 >> @@ -0,0 +1 @@ >> +SQL error: descriptor "mydesc" not found on line 31
> does not look like the same kind of failure as what we've been dealing > with up to now. So maybe what we've got is that we fixed the stdio > loss problem, and now the error rate is down to the point where we can > notice other, even-lower-probability issues. Yeah, I think so. I grepped the buildfarm logs for similar failures and found three occurrences: https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=bowerbird&dt=2019-02-03%2018%3A36%3A05 https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=bowerbird&dt=2019-01-17%2014%3A30%3A07 https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=drongo&dt=2020-01-02%2018%3A03%3A52 All of these are in the thread/descriptor test, failing at the deallocate step in for (i = 1; i <= REPEATS; ++i) { EXEC SQL ALLOCATE DESCRIPTOR mydesc; EXEC SQL DEALLOCATE DESCRIPTOR mydesc; } where the test is running several of these in different threads. I wonder whether there's some missing thread-locking in the ECPG descriptor support. It is odd though that we have seen this only on Windows members. Low-probability or not, you'd think we'd have some similar reports from non-Windows critters if it were possible. regards, tom lane
