On Wed, Oct 11, 2017 at 11:15:57PM +0100, Ramsay Jones wrote:
> Hi Adam,
>
> I had a test failure on the v2.15.0-rc1 build tonight.
> The test in question being t0021-conversion.sh #15
> ('required process filter should filter data'). I didn't
> have any test failures on v2.15.0-rc0, and I don't see
> any change that would have affected this test.
>
> Also, I ran this test by hand (well, in a shell loop) at
> least 70 times tonight (after the test-suite run), without
> any failures, so ... (unfortunately, I don't have a trash
> directory to look at. :( )
>
> I have just kicked off another full test-suite run.
>
> Just a heads up! ;-)
Hi Ramsay,
I assume, given you're emailing me, that this is a Cygwin failure?
t0021.15 has PERL as a requirement, and I see semi-regular failures from
Git tests that are Perl-based in one way or another (git-svn tests are
the most common problems). I've not spotted t0021 failing in that way,
but it sounds like the same class of problem.
I dig into these failures when I see them, mostly by running the script
a few hundred times until I get the failure again, and they've always
been Perl itself segfaulting. That points to the problem being in
Cygwin's Perl package rather than Git, and it's very unlikely to be
anything that's got worse in v2.15.0.
I've never managed to get further than pinning the blame on Perl,
though. If you manage to reproduce the failure and it turns out to be
something different, or you manage to dig into the failure in Perl
itself, that'd be some very interesting news.
Adam
