On Sun, Jan 06, 2019 at 08:28:33PM +0100, Paul Gevers wrote: > Hi Adrian, Niels,
Hi Paul, > On 06-01-2019 20:01, Adrian Bunk wrote: > >> That is because gmsh from testing links to libmed1v5. Adding this > >> *versioned* breaks to libmed11 (albeit being a bit ridiculous from the > >> archive point of view) would do the right thing AFAICT. > >> ... > > > > despite libmed11 not being installed at all in the debci test? > > britney reads this to determine which packages (based on SRC) need to > come from unstable for testing in testing. > > > You are saying that the (test) dependencies of gmsh/unstable matter when > > testing gmsh/testing with med-fichier/unstable? > > Yes, because they are processed by britney. So if gmsh/testing and > med-fichier/unstable aren't a good match based on information that > britney has available, it will request the test with both from unstable. > britney looks at the regular Depends/Breaks/Conflicts fields as well as > the test dependencies. > > > That's unexpected compared to the normal dependency semantics. > > I agree, but this is how it is done now until we change it. We're open > for suggestions and I assume the release team is open for patches as well. when there is a reason for a maintainer to declare that a new upload breaks the autopkgtest only of a reverse dependency, the logical way would be someting like X-Breaks-Autopkgtest-Of: gmsh (<< 3.0.6+dfsg1-4+b1) This would be a proper interface for maintainers that doesn't expose whatever black magic britney+debci are doing. > >>> The root problem is that debci installs cruft packages from unstable. > >> > >> Care to elaborate what you mean here? debci doesn't install anything. > >> It's apt that installs stuff. Based on a slightly odd configuration put > >> in place by autopkgtest on request of debci which got its trigger from > >> britney. > > > > Britney says for med-fichier: > > old binaries left on amd64: libmed1v5, libmedc1v5 (from 4.0.0+repack-1) > > (but ignoring cruft, so nevermind) > > > > Installing one of these cruft packages that cannot ever migrate to > > testing is the problem. > > Sure, but the root cause is that the combination to test isn't properly > determined by britney. britney just doesn't know. Britney says in excuses that the package is cruft. > > Correct would be that this debci test does not pull in a single package > > from unstable, since no non-cruft package depended on from gmsh/testing > > is being provided by med-fichier/unstable. > > I don't agree. What we want to test is what happens if med-fichier has > performed the transition from libmed5v1 to libmed11 and we try to move > the set to testing. Note that with smooth updates this is not a "set" - it does happen that a so-name changing library package migrates to testing before any of the reverse dependencies has been rebuilt. With the new de facto default of 2 days for migrations it is not even rare. The med-fichier binNMUs were 3 days after med-fichier was uploaded. Without this debci issue med-fichier would have already migrated before gmsh was binNMUed in unstable. > Paul cu Adrian -- "Is there not promise of rain?" Ling Tan asked suddenly out of the darkness. There had been need of rain for many days. "Only a promise," Lao Er said. Pearl S. Buck - Dragon Seed