Archive skew and multiarch packages

[Scott Ritchie]

As beta approaches I'm facing a bit of a dilemma.  Originally I had the
following package relationship on amd64:

wine1.4
 - depends: wine1.4-i386 (a package only built on i386 and M-A: foreign)

I discovered a package in the archive (lmms) that actually builds with
64 bit wine1.4, which means the above relationship is not satisfiable on
the build daemons.  So I changed it to this:

wine1.4
 - depends: wine1.4-i386 | wine1.4-dev

Somewhat naively thinking this would make the build daemons happy (it
does), but that most users would get the default wine1.4-i386.


However, wine1.4-i386 has about 18 dependencies, plus itself, and each
and every one of them needs to be exactly the same version as their
amd64 version in order to be installable on a multiarch system.

This means that if a user installs or I push an update during a time
when one of those 18 packages just happens to be taking longer to build
on i386 than amd64 (or vice-versa) then apt will give up on
wine1.4-i386, install wine1.4-dev, and they will be stuck with a
(near-useless) 64-bit only wine.


I'm tempted to just break lmms' build and revert the change, going with
a hard dependency on wine1.4-i386, but it feels like this sort of issue
is going to come up again.  There are other reasons to do this - for
example 64-bit only Wine has no real way of detecting a 32-bit
application and prompting the user to install the proper package.  If a
true needs-64-bit-wine to build application creeps its way into Precise
post release, then I can always issue an SRU (which are immune to
archive skew).

Regardless, I'd like to discuss this sort of thing at UDS.

Thanks,
Scott Ritchie