Re: [gentoo-project] Call for agenda items -- Council meeting 09-10-2012

[Brian Harring]

On Thu, Sep 27, 2012 at 08:19:13AM +0200, Ulrich Mueller wrote:
> >>>>> On Tue, 25 Sep 2012, Ulrich Mueller wrote:
>
> > - Package names:
> >   Our current spec forbids that package names "end in a hyphen
> >   followed by one or more digits". This isn't consequent, since it
> >   still allows PN to be e.g. "foo-1a" which looks like a valid PF.
> >   OTOH, there's no technical reason for this limitation (backwards
> >   compatibility issues taken aside).
> >   Since this issue is open since more than five years, I believe that
> >   it's time to ask the council for guidance in what direction we
> >   should go:
> >     a) Drop the limitation entirely (possibly in a future EAPI).
> >     b) Make it stricter, i.e. disallow package names ending in a
> >        hyphen followed by anything that looks like a valid PVR.
> >        This is current Portage behaviour, and the tree complies with
> >        it, too.
> >     c) Leave the spec as it is (and make Portage comply with it).
> >   See bug 174536 for details.
>
> Actually, there's also:
>     d) Require a) for Package managers and b) by tree policy
>        (Postel's Law, brought up by mgorny). Practically, this would
>        mean that repoman would reject "foo-1" as package name, but the
>        rest of Portage would accept it.

Grr.  And people wonder why I get brutally blunt at times.

This restriction, whether people like it or not, is encoded into 
some fairly core places that have consequences.  You do this, you 
break all existing managers that see it in the vdb.  Not "oh looky, a 
warning, isn't that cute by mildly annoying", break.

Badly. Break.  PM shats the bed, traceback gets thrown.

Considering managers do full graphs, that node gets touched, the 
affected PM isn't going to be able to even upgrade it's way out of it- 
best case, if it's implemented sanely a --nodeps will avoid that node.  
However a lot of PMs still have legacy virtuals code, meaning if that 
cache is stale, looky looky, the vdb gets walked and you get a 'boom'.

This isn't even commenting on binpkgs, which is where you're likely to 
see older managers existing as clients.

All of that is unversioned.  We cannot hide this sort of change- at 
best we can deploy it, force all QA tools to block it for N years, 
than allow it in; anything else is risking known (yes known, people 
have been pointing this out for ages).

What for?  So someone can name their package foo-1?  Is that really 
such a major gap we're willing to induce breakage?

Will anyone even !@#*ing use a package names foo-1?  I've yet to see 
an example given, just ignoring of the breakage it will induce.

Honestly, you want to push it, you address how this isn't going to 
break things, lay out the timelines, identifying when we willingly 
switchover to breaking managers older than a certain date.  It does 
not get pushed up to the council w/out the negatives/flaws mentioned.

Shorter version; you very clearly left out option C; "leave it as is 
since PMS is filled with warts, this isn't hurting anything, and 
changing it will break things."

~harring