Dnia 2015-04-06, o godz. 02:38:41
Andrew Savchenko <bircoph@gentoo.org> napisał(a):

> On Sun, 5 Apr 2015 18:54:10 -0400 Rich Freeman wrote:
> > On Sun, Apr 5, 2015 at 5:27 PM, Andrew Savchenko <bircoph@gentoo.org> wrote:
> > >
> > > 2. Otherwise allow developers to drop stable keywords from affected
> > > package and _all_ its reverse dependencies. This way a part of
> > > stable tree will be removed, but only a part. With this approach
> > > arch teams will be freed of an extra burden, while they will be
> > > still able to maintain a smaller stable tree.
> > >
> > > This is a win-win solution: a stable tree will be still kept in a
> > > maintainable size and developers will not have a long-term blockers
> > > on their stabilization requests.
> > >
> > > 3. And last but not the least: apply the rules above to all arches,
> > > not just minor teams (though probability that amd64/x86 will be
> > > slow is lower, of course).
> > >
> > 
> > This was some of what I was getting at.  My question still stands that
> > I'm not sure arch teams REALLY want 300 packages to have their stable
> > keywords removed instead of just having one package break the
> > depgraph.
> 
> Hmm, that's a hard question. I tried to consider this issues from a
> point of view of a user of such arch. If package is not used or
> user may delete it and its deps without much harm, it doesn't affect
> user at all. If it is used and needed, then in case of:
> 
> - one package with removed stable keyword a user have to add to
> package.keywords only a single package, though it might be
> difficult to locate such package, because portage deptree failure
> events may be really obscure sometimes;
> 
> - all subtree of stable keywords is removed; then user have to
> add all these packages to package.keywords, portage messages should
> be clear here (but one never knows), though manual keywording of
> hundred of packages will be irritating at best (even using "cat/*"
> masks). So if number of affected installed packages is large, users
> will likely move to ~arch all their setup.
> 
> So from user's perspective stable deptree broken in single point is
> a better solution, but(!) if portage will cleanly suggest this
> point.
> 
> Another issue to consider: what if we have one such package that
> broke stable deptree, then after awhile another one and so on. In
> the result stable tree will got corrupted beyond repair.
> 
> Maybe some grace period will help here? E.g. remove stable keyword
> from a single package, wait for 30 days (or so) for reaction from a
> team, and then dekeyword all reverse dependencies.

Which means developers can't commit properly for 30 days. Really
awesome solution, thank you. Please ping QA instantly once you do it,
you'll save us time figuring who to ban for the breakage.

Let me remind you: once you break dependency tree on *ANY* stable
architecture, repoman won't let you commit. Travis turns red. All pull
requests are marked as broken. Long story short, we end up with a lot
of extra work.

And so far, nobody but me and Patrick basically cared about dependency
graph not being broken.

-- 
Best regards,
Michał Górny