Push scripts, mash (was: Re: FESCo wants to ban direct stable pushes in Bodhi (urgent call for feedback))

[Michael Schwendt]

On Fri, 05 Mar 2010 11:03:12 +0100, Kevin wrote:

> Yeah, basically "mash" is a really brute force solution, I think directly 
> writing out only the new updates as the first prototypes of Bodhi did and as 
> the Extras scripts also did/do is a much smarter solution. Always 
> recomputing everything sucks.

It can be beneficial. Also the tagging in koji adds possibilities that are
nice. Such as optional reconstruction of repositories with builds stored
elsewhere.
However, if there are resource constraints, such as slow NFS access
to the packages and repos, rebuilding less sounds more appropriate.

> It was claimed that recomputing is necessary for some obscure multilib 
> corner cases. 

What might those be?

I could imagine "obsolete multilib packages". They have been pulled in
previously and later become unnecessary for an update and its changed
dep-chain. Old RepoPrune kills them whenever the src.rpm is updated for
the primary arch (=> old multilib pkgs would refer to a src.rpm that
isn't the newest one => prune those). And recomputing the full multilib
dep-chain on demand could still be possible. Deleting packages, which
have been published before, is dangerous, though. Users may have
installed those packages (even if just as automatic dependencies).

> Let me suggest a radical solution for that: drop multilib repos!

Or return to white-listing only a small selection of packages.

There are just too many -devel packages and their dependencies to be ever
relevant to someone for multi-arch installs. Far more users install i686 on
64-bit CPUs, and I have doubts that x86_64 installation users do much
development with i686 packages. At most they install 32-bit apps where
64-bit builds aren't available or "less good".