Bug#923238: libmarc-charset-perl: needs a rebuild on 32bit architectures?

[Niko Tyni]

On Wed, Feb 27, 2019 at 09:20:25PM +0200, Niko Tyni wrote:
> On Mon, Feb 25, 2019 at 11:31:14AM +0100, Gianfranco Costamagna wrote:
> > Package: libmarc-charset-perl
> > Version: 1.35-2
> > Severity: serious
> > 
> > Hello, for some reasons the package testsuite started to fail in Ubuntu for this package and xml-perl reverse-dependency,
> > only on armhf and i386.
> > This happened when the new gdbm has been uploaded and rebuilds issued.
> > 
> > I traced down the problem to some differences in the march8/utf8 Table generation, I don't know how serious it is, but the
> > testsuite seems completely broken on armhf and i386 at least, and utf8 cjk conversion seems to return wrong values.
> > This is the reason for me opening this bug as "serious".
> 
> Thanks for noticing this. I've confirmed that this happens on at least
> Debian sid/i386 too. It's a bit unfortunate that we only have autopkgtest
> checks on amd64, so this wasn't spotted earlier.
> 
> > after a no-change rebuild of the package, and installing it, the test goes passing ok:

This was discussed in #923609, and src:gdbm will not restore backward
binary compatibility. The incompatibility is due to building with
LFS support in the newer versions. The gdbmtool package now includes
separate -nolfs binaries compatible with old databases, so that they
can be converted to a compatible binary format.

This means that libmarc-charset-perl indeed needs to be rebuilt.  However,
a simple binNMU does not do anything to prevent broken combinations of
libmarc-charset-perl and perl in partial upgrades.

I suggest the following steps to fix this:

- have perl_5.28.1-5 Build-Depend on libgdbm-dev (>= 1.18-3)
  [any gdbm version built with LFS support would do, but 1.18-3
   fixed other binary compat issues so pick that for safety]

- have perl_5.28.1-5 Break libmarc-charset-perl (<< 1.35-3)

- have libmarc-charset-perl_1.35-3 Build-Depend and Depend on
  perl (>= perl_5.28.1-5)

We could limit the above to just 32-bit architectures, but IIRC those
would need to be listed one by one and it's probably not worth the
trouble.

This will cause temporary uninstallability of libmarc-charset-perl in
sid so the uploads should be coordinated a bit. I guess I can do both
if needed.

Will clone a separate bug against perl soon but would appreciate some
eyeballs first.
-- 
Niko