[Yum-devel] RFC: Best providers repo. metadata (initial implementation)

[James Antill]

 There are a growing number of cases where we have a virtual provide X,
that is satisfied by N packages but the repository owner knows that only
N-M should be considered.
 For instance:

1. smtpd, smtpdaemon, /usr/sbin/sendmail, MTA, server(smtp)

   These are provided by a lot of things, where Fedora/etc. will almost
certainly only want postfix installed instead of exim (shortest name).

2. java-devel

   Here we actually have had bug reports in Fedora (not the only ones),
as openjdk and gcj provide the same thing, and again shortest name wins
but Fedora wants openjdk to win.

 So here's my initial implementation:

1. Create a simple XML format for "best providers", idea being that it
won't be that big and only needs provides and package names. Sample file
is:

<?xml version="1.0" ?>
<bestproviders>
  <bestprovider>
    <provide>/usr/sbin/sendmail</provide> <provide>MTA</provide>
    <pkgname>postfix</pkgname>
  </bestprovider>
</bestproviders>

2. Yum looks for a 'best_providers' metadata type from repomd.xml and
creates a per. repo. 'best providers' followed by a global one.

3. If we hit compare_providers() then we filter to data in the best
providers and give them a big score increase.

4. I also changed "yum install smtpd" to go through compare_providers()
instead of it's own simplified version it had.


 Patch is at:

http://james.fedorapeople.org/yum/patches/yum-best-providers-metadata.patch

 Sample repo. with best_providers data (i386 and x86_64 only):

[and.org-OS-updates]
name=And.org OS updates
baseurl=http://www.and.org/yum/fc$releasever-$basearch-updates
enabled=true


 Comments? ... things to think about:

1. It's only provide => package name ... is that enough? Possibility of
more data: allow full nevra type preferences, allow conditions based on
what/if package requires the provide and allow specific scores so pkgA
is preferred over pkgB, etc.

2. Atm. we say "if the requiring package comes from repoA, then do a
best provider lookup in repoA ... if that fails do a global lookup. If
we don't have a requiring package we just do a global. For the global
data we just merge everything from all repos.
 Is this too much? My intent was to allow two repos. to provide say
my-video, and one to prefer gstreamer but one prefer xine, but I'm not
100% sure it's worth it ... it's certainly easier to just have a global.
 Is this not enough? As above, but I could see cases where you'd want to
say "use Fedora, but with this local repo. and have BLAH as the
preference for FOO".

-- 
James Antill <james at fedoraproject.org>
Fedora