Re: [PATCH] add the ability to create and retrieve delta objects

From: Nicolas Pitre <nico@cam.org>
Date: May 5, 2005, 03:25 UTC
Message-ID: <Pine.LNX.4.62.0505042311550.14033@localhost.localdomain>
In-Reply-To: <Pine.LNX.4.58.0505041501220.2328@ppc970.osdl.org>

On Wed, 4 May 2005, Linus Torvalds wrote:

> I'll happily take the patch and just not use the delta packing myself (at 
> least until I trust it). But before I take the patch I want to make sure 
> that people agree on it, and that it's been tested well enough that it 
> won't cause people to corrupt their repositories.

To that effect I'm adding knowledge of delta objects to fsck-cache, to verify lists of deltas are all reachable and that they all expand to the expected data. This way it would be a good test to completely deltafy a whole repository, run fsck-cache to ensure everything is fine, and undeltafy it all to see if things are still sane.

> For example, I do _not_ want to be in the situation SVN is in, where if
> you corrupt your SVN database, you're totally screwed. There's a real
> advantage to not having fancy data structures or complicated consistency
> rules.

With deltas it is still a bit more risky by design since many objects end up depending on a single one. You loose that top object and it's all the delta chain that's gone. But having the choice to use them or not is what makes the whole system flexible and suited to anyone's balance between robustness vs disk space. Converting back and forth is certainly not a problem with the git model.

And if you deltafy things such that the objects in your head tree are always top of delta chain then you're not badly affected if some intermediate delta objects are corrupted since they are part of old trees only.

Nicolas