Bug#1005298: libdevel-nytprof-perl: autopkgtest failure with Perl 5.34: t/test62-subcaller1-b.t

[Niko Tyni]

Package: libdevel-nytprof-perl
Version: 6.11+dfsg-2
Severity: serious
User: debian-perl at lists.debian.org
Usertags: autopkgtest

This package fails its autopkgtest checks on current sid, triggered by
Perl 5.34.

#   Failed test 'test62-subcaller1-b.calls match generated calls data for blocks=1 calls=1 compress=1 leave=1 savesrc=0 slowops=2 start=init use_db_sub=0'
#   at t/lib/NYTProfTest.pm line 468.

The test suite passes at build time, so this wasn't caught in our
pre-transition testing.

The failing tests have pregenerated traces of expected execution details.
These traces changed during the 5.34 cycle, so there are two versions
of this test: test62-subcaller1-a.t for < 5.34 and test62-subcaller1-b.t
for 5.34 onwards.

The new execution trace is unfortunately sensitive to the test
environment: if the binary plugin (NYTProf.so ) gets loaded by
Devel::NYTProf::Core from a relative filesystem path (like during the
build, from under blib/), the execution trace is different than when
NYTProf.so gets loaded from an absolute path (like during autopkgtest
checks, from the system directory.)

The specific difference seems to be that XSLoader.pm
will fall back to DynaLoader for relative paths around

 https://sources.debian.org/src/perl/5.34.0-3/dist/XSLoader/XSLoader_pm.PL/#L128

I haven't investigated this further.

So nothing is broken here, the test just works only during build.
I can't see a clean fix, and I don't think we should bother upstream
with these details. I suggest just skipping this test during autopkgtest.

(Testing  against installed versions is a Debian-specific thing that CPAN
distributions aren't designed for, it just happens to work almost always.)

(Since this is a test-only thing, we probably shouldn't make the new perl
 Break unfixed versions.)

See also https://github.com/timbunce/devel-nytprof/issues/143 about
the fragility of the tests.
-- 
Niko Tyni   ntyni at debian.org