Re: inconsistant regression test results...

Vikram Kulkarni <vkulkarn(at)brownforces(dot)org> writes:
> I'm trying to build+install Postgresql 7.2.1 on a OpenBSD 3.1-stable
> computer. The first time I built it, the 12/79 of the regression tests
> failed. This scared me, so I did a gmake distclean and then reconfigured
> and rebuilt everything. This time, 14/79 tests failed. ...
> This time, 11/79 test failed. This got me wondering, so I reran the
> entire process (untaring, configuring, gmake'ing, and gmake check'ing)
> three more times. Different results each time (14, 15, then 10).

It looks to me like the primary failures are that tests abort with
either
psql: Server process fork() failed: Resource temporarily unavailable
or
psql: could not send SSL negotiation packet: Broken pipe

Some later tests may then fail because they expect to find tables or
data created by the un-executed earlier tests.

The fork-failed messages suggest very strongly that you are running out
of kernel resources when you get more than a dozen or so server
processes going. Perhaps you are too low on swap space, or need to
enlarge the kernel's file table size. You could try to confirm this
by running the regression tests serially instead of in parallel (use
the installcheck option); or you could modify the parallel_schedule
file to break apart the more highly parallel test sets into smaller
groups. If the tests pass that way then the problem is triggered by
load, not by any specific test.

Not sure about the SSL complaint, but I suspect it's the same problem at
bottom. You should look in the postmaster log file that's generated by
the make check run, and see if you can find what gets logged by the
postmaster when one of those failures is seen on the client side.

regards, tom lane