Re: WARNING: THIS PATCH CAN BREAK YOUR REPO, was Re: [PATCH 2/3] Only repack active packs by skipping over kept packs.

From: Shawn Pearce <spearce@spearce.org>
Date: Oct 30, 2006, 21:09 UTC
Message-ID: <20061030210908.GB5775@spearce.org>
In-Reply-To: <20061030205200.GA20236@delft.aura.cs.cmu.edu>

Jan Harkes <jaharkes@cs.cmu.edu> wrote:

Show 19 quoted lines

> On Mon, Oct 30, 2006 at 03:26:11PM -0500, Shawn Pearce wrote:
> > Actually the breakage is easier to reproduce without trashing
> > a repository.
> > 
> > Do the above so you have everything in one pack.  Now use rev-list
> > to simulate the object list construction in pack-objects as though
> > we were doing a 'git repack -a -d':
> > 
> >   git-rev-list --objects --all \
> >     --unpacked=.git/objects/pack/pack-*.pack \
> > 	| wc -l
> > 
> > gives me 102 (WRONG WRONG WRONG WRONG!!!!!!)
> 
> The problem seems to be that as soon as we hit something that is found
> in a pack that is not on the ignore list, that object and all it's
> parents are marked as uninteresting. So if the kept pack contains a
> slice of commits (v1.4.3..v1.4.3.3) the revision walker will only return
> the recent stuff (v1.4.3.3..) and drop the older data (..v1.4.3).

Right - I got that far in my own research and then saw your patch
drop into my inbox. :-)

> The following patch does fix the problem Nicolas reported, but for some
> reason I'm still getting only 102 objects (only tags and the commits
> they refer to?) with your test.

Ack'd.

Your patch fixes both bugs for me. The rev-list test I talked about above is now turning up 31846 objects which is the correct count. The repack test Nico crafted works correctly too.