Re: cloning a namespace downloads all the objects

From: Jeff King <peff@peff.net>
Date: Nov 16, 2012, 17:34 UTC
Message-ID: <20121116173416.GA13268@sigill.intra.peff.net>
In-Reply-To: <7vvcd57b32.fsf@alter.siamese.dyndns.org>

On Fri, Nov 16, 2012 at 08:43:13AM -0800, Junio C Hamano wrote:

Show 21 quoted lines

> > I have a repo on a server, which contains, as namespaces, the contents
> > of several different repos of varying sizes.  When I run a clone
> > command for the smallest of the namespaces (I have a script that
> > intercepts the clone and sets GIT_NAMESPACE appropriately), I get the
> > correct set of refs, but *all* the objects from *all* the namespaces.
> >
> > And since no refs from the other namespaces have come down, a 'git gc
> > --prune=now', run immediately after, reduces the size of
> > ".git/objects" to the size I would expect for just that small
> > namespace.
> >
> > In effect, it is bringing down data that is not reachable and will be
> > wiped out on the next gc.
> >
> > Is this expected?
> 
> I do not think so.
> 
> This was done with a series between a1bea2c (ref namespaces:
> infrastructure, 2011-07-05) and bf7930c (ref namespaces: tests,
> 2011-07-21); Josh, care to comment on and to look into it?

I'd guess that the "create_full_pack" logic in create_pack_file is to blame. The client asked for everything we advertised, so we pass "--all" to pack-objects rather than giving it the specific list of tips.

We'd have to either fix that logic, or teach the pack-objects subprocess to respect GIT_NAMESPACE when processing --all.

-Peff