[Pkg-julia-devel] Bug#816980: Bug#816980: julia: FTBFS when built with dpkg-buildpackage -A (signal (11): Segmentation fault)

[Peter Colberg]

On Sat, Mar 12, 2016 at 10:44:44AM +0100, Santiago Vila wrote:
> On Fri, Mar 11, 2016 at 06:29:40PM -0500, Peter Colberg wrote:
> > Could you retry by building julia 0.4.3-3 from experimental?
> 
> I'm only building packages on testing, sorry.

Ok, then please retry once julia 0.4.3-4 has been uploaded to unstable
and propagated to testing. The above versions use LLVM 3.8 instead of
LLVM 3.7, which improves test run time significantly, and might also
reduce memory usage of individual tests.

> > It builds fine on my amd64 machine using "sbuild -j4".
> 
> That's fine, but did you actually test "dpkg-buildpackage -A"
> in stretch?

Yes, I forgot to mention that I also successfully tested

sbuild -j4 --debbuildopts=-A

In fact, the arch-indep build was fixed in julia 0.3.12-2.

> So why do we need to run the test suite for arch-indep builds?
> 
> Version 0.3.11-1 took only 2 minutes to build (with "dpkg-buildpackage -A")
> and it generated the arch-independent packages without any problem.
> 
> Do we really need a full build to generate the arch-independent
> packages? (This was not needed in the past, apparently).

Yes; at minimum, the arch-indep build needs to generate some files
that are included in the package julia-common, which was added in
version 0.3.12-1 to address a lintian warning about shipping to
many common files in the arch-dependent package julia.

The tests ensure that the files installed to julia-common are usable.
We could drop these tests once autopkgtest (debian/tests/control) has
been integrated into the Debian infrastructure such that it prevents
packages with failing autopkgtest from entering the archive.

> > Does your machine have enough memory?
> 
> I hope so. My machine is a qemu virtual machine with 5GB of RAM and
> 2GB of swap. Is that not enough?
> 
> Moreover, it has only one CPU and AFAIK parallel build should only be
> enabled when explicitly requested.

This is the case, please see the variable TESTS_ENV in debian/rules.

Julia requires at least two processes for testing, which at 500 MiB
per worker requires a total of 1GB, or 2GB to be safe. (The workers
are restarted only after they hit the memory limit, which means
individual tests may consume more then JULIA_TEST_MAXRSS_MB.)

> My main problem when testing packages with "dpkg-buildpackage -A" is
> that we don't have a "Build-Memory:" control field yet.
> 
> What would be the value of this field in "julia" if such field existed?
> (For the amd64 architecture and only one CPU).

Such a field is difficult to define; as described above, for julia the
value depends on the number of parallel processes, and even then the
exact figure is hard to determine.

Please let us know how julia 0.4.3-4 fares on your machine once in testing.

Regards,
Peter