[Mageia-sysadm] Packages SVN

[Luca Berra]

On Wed, Nov 03, 2010 at 11:12:53PM +0100, Olivier Blin wrote:
>nicolas vigier <boklm at mars-attacks.org> writes:
>
>> We can also remove history on cauldron binary branch, although it will
>> require to stop new commits for a few minutes/hours, doing something like
>> this :
>> http://robmayhew.com/delete-parts-of-subversion-history/
>
>Right, we could do in such a way.
>Each N distro branch could contain binary history from N-1 to N.
>
>For example when we release 2011.1, we would have binary history in the
>2011.1 distro branch from 2011.0 to 2011.1, and start cauldron from a
>one-revision copy.
>
>Though, there is a disadvantage that could get problematic, when we do
>backports, we would have to copy the binaries to all target distros, and
>it would not be a cheap SVN copy anymore.
>This could get pretty big with games and OOo...

I wonder why use svn at all in order to store blobs
IMHO it does not give any advantage
source tarballs in packages are usually compressed, so no delta is
possible, besides that sources are usually removed and the new one is
added, so it wont be even tried.
removing something from SVN is a PITA
creating a branch for each distro will waste enormous amounts of space,
and will create the issue pointed above by blino.

why not just use flat files to store flat files?
iirc redhat is doing this to store files

if i were to design this i would index flat files by sha
and modify repsys/mdvsys in order to commit a file containing the sha in
svn://specrepo/package/SOURCES/foo-0.0.1.tar.xz instead of the real blob
and checkout the real file from file://binrepo/hashed_dir/sha instead of
the dummy.

Cleanup can be done programmatically by retrieving dummy files from the
specrepo, with a set criteria (eg stuff tagged with a distro release
number), and killing the unreferenced ones.

This would also be useful spacewise, because we could safely kill blobs
that were never released: when in cauldron we upload a svn/git version
of foo to test, but in the release a real version is included. After a
while killing all the foo-0.0.1.svn203892 that never went into a release
would be nice and noone will miss them.

It would also allow to share the same blob beetween two different
packages, maybe less useful, but it would come for free.

L.

-- 
Luca Berra -- bluca at vodka.it