On Mon, Jul 03, 2017 at 10:40:45PM +0300, Niko Tyni wrote: > Package: autopkgtest > Version: 4.4 > User: debian-p...@lists.debian.org > Usertags: versioned-provides > X-Debbugs-Cc: debian-p...@lists.debian.org > > As seen on ci.debian.net with for instance libhttp-tiny-perl and > libcpan-meta-perl, autopkgtest gets confused about versioned Provides > that were introduced in sid recently with perl_5.24.1-5. > > It looks like "Depends: @" will no longer pull in the binary packages > to be tested if the same name is also Provided by installed packages > with a version. > > My reading of the autopkgtest code is that it synthesizes a dependency > on 'package (>= 0~)', where the versioning is assumed to guarantee that > only a real package gets pulled in. This assumption no longer holds with > versioned Provides.
Oh, I'd mostly forgotten about #761003 which introduced the '(>= 0~)' thing three years ago and where we predicted that this will break with versioned Provides. The solution we discussed there was to insert an additional explicit apt-get install <package> phase to the autopkgtest-satdep.deb installation (which unpacks a temporary package and then calls 'apt-get --fix-missing' to have apt solve the dependencies.) This explicit call prefers real packages to virtual ones (at least in all my tests though I can't find this documented; possibly this should be checked with the apt maintainers.) Is this approach something you would consider now that the versioned Provides issue has materialized in practice? -- Niko Tyni nt...@debian.org
