Mattia Rizzolo <mat...@debian.org> writes: > On Sat, Jan 05, 2019 at 04:02:35PM +0100, Ole Streicher wrote: >> I'll do tonight. It however looks a bit suboptimal: when the CI test >> with a new version fails for an old reverse dependency, then the new >> version obviously breaks that old package. So, the breakage could be >> detected (and taken into account) automatically. Why do we need a manual >> action then? > > Let me try to suggest you see from another side: the goal of the Breaks > field is to prevent a given combination of known broken packages to be > installed. > Currently autopkgtest blocks the migration of numpy, but if there wasn't > that block a Breaks should still be added, as the astropy in testing is > not compatible with the new numpy. The fact that it hints britney to > trigger the right tests is just a "side effects", the Breaks should be > added nonetheless, theoretically; in practice, we rarely did it before > autopkgtest started blocking the testing migration.
I agree (with the explanation in your other mail) here; it just states the fact that they really don't go together very well (at least in situation similar to the specific CI test). >> Python-astropy is however going to be removed completely; it has >> however some cruft rdeps left in unstable. So, it cannot removed from >> unstable now, and therefore still remains in testing and >> (unnecessarily) blocks the numpy migration. Python-astropy already >> has an RC bug, but its autoremoval from testing is still not even >> announced yet. > > Maybe you could ask the release team to hasten the removal of > python-astropy and rdeps from testing? If the plan is to not relase > it in buster I don't see a reason to keep it further. Couldn't we just also add a "breaks" to numpy? The important fact here is, that the new numpy and the current python-astropy don't work together; and this is independent of whether a fixed python-astropy version exists. This would remove one dependent party (release team) from the chain of blocking causes for the migration. >> The migration blocking CI tests seem to cause much more headaches than >> just "fix your bugs"... > > Well, from what I see, it has helped a lot in this half a year detect a > ton of regressions that would have otherwise bothered several people in > a later moment… I usually regularly look into my packages and file bugs if a CI test fails caused by a buggy updated dependency. This works well without the need of blocks. It would also work, when a failing CI test would automatically trigger a bug report against the updated package, which then could be re-assigned to the rdep in case the problem was there. So, I value very much the CI tests by themself: they are the greatest invention since sliced bread, but I still dislike the inflexible handling of blocks. Often, a failing CI test makes a package only buggy with respect to a certain feature, so it would correspond to a "important" or "normal" bug severity. They are however always handled like a "serious" bug by preventing migration. I miss the "That was not a serious failure" button. Thanks fur your help! Best Ole
