Re: Hosting Git repositories: how useful will git-gc be?

From: Jeff King <peff@peff.net>
Date: Sep 5, 2009, 06:24 UTC
Message-ID: <20090905062424.GC29863@coredump.intra.peff.net>
In-Reply-To: <vpq1vmo9xai.fsf@bauges.imag.fr>

On Thu, Sep 03, 2009 at 11:45:25AM +0200, Matthieu Moy wrote:

> A question: is it necessary/recommanded/useless to set up a cron job
> doing a "git gc" in each repository? My understanding is that a push
> through ssh will do some packing, is it correct? Does receiving a pack
> trigger a "git gc --auto"?

The objects are transferred as a pack. If the number of objects is less than receive.unpackLimit (default 100), then they are unpacked to loose objects. If more, we keep the pack, after completing any missing deltas used by a thin pack.

So if you tend to push frequently, you will end up with a lot of loose objects. Even if you have packs, they will be larger than necessary because you will be missing deltas between objects across packs. And of course you will eventually end up with a large number of packs, which is less efficient (each pack has an index, but I believe we search the indices linearly).

Receiving a pack does not (AFAICT looking at the code) trigger a "gc --auto". Running it has other benefits, too, like pruning cruft and packing refs. So I think it is probably a good idea to run it periodically.

Running it daily or weekly is probably reasonable. You could run it on every push using the post-update hook, but that may cause excessive I/O for very little benefit.

-Peff