Re: Errors cloning large repo

From: Shawn O. Pearce <spearce@spearce.org>
Date: Mar 11, 2007, 02:00 UTC
Message-ID: <20070311020013.GD10343@spearce.org>
In-Reply-To: <esu11i$m54$1@sea.gmane.org>

Jakub Narebski <jnareb@gmail.com> wrote:

Show 12 quoted lines

> Shawn O. Pearce wrote:
> 
> > One thing that you could do is segment the repository into multiple
> > packfiles yourself, and then clone using rsync or http, rather than
> > using the native Git protocol.
> 
> By the way, it would be nice to have talked about fetch / clone
> support for sending (and creating) _multiple_ pack files. Beside
> the situation where we must use more than one packfile because
> of size limits, it would also help clone as it could send existing
> packs and pack only loose objects (trading perhaps some bandwidth
> with CPU load on the server; think kernel.org).

I've thought about adding that type of protocol extension on more than one occasion, but have now convinced myself that it is completely unnecessary. Well at least until a project has more than 2^32-1 objects anyway.

The reason is we can send any size packfile over the network; there is no index sent so there is no limit on how much data we transfer. We could easily just dump all existing packfiles as-is (just clip the header/footers and generate our own for the entire stream) and then send the loose objects on the end.

The client could easily segment that into multiple packfiles locally using two rules:

  - if the last object was not a OBJ_COMMIT and this object is
  an OBJ_COMMIT, start a new packfile with this object.

  - if adding this object to the current packfile exceeds my local
  filesize threshold, start a new packfile.

The first rule works because we sort objects by type, and commits appear at the front of a packfile. So if you see a non-commit followed by a commit, that's the packfile boundary that the server had.

The second rule is just common sense. But I'm not sure the first rule is even worthwhile; the server's packfile boundaries have no real interest for the client.

But Linus has already pointed all of this out (more or less) in a different fork of this thread. ;-)

-- 
Shawn.