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

From: Chris Mason <mason@suse.com>
Date: May 3, 2005, 16:09 UTC
Message-ID: <200505031209.52460.mason@suse.com>
In-Reply-To: <Pine.LNX.4.58.0505030804170.3594@ppc970.osdl.org>

On Tuesday 03 May 2005 11:07, Linus Torvalds wrote:

Show 6 quoted lines

> On Tue, 3 May 2005, Chris Mason wrote:
> > On the full import of all the bk->cvs changesets, the average file size
> > in .git is 4074 bytes.  73% of the files are 4096 bytes or smaller.
>
> Have you checked how many of those are blobs?
>

I've got cg-admin-lsobj running (effectively find .git -type f | xargs cat-file), it is taking a looong time but the ratios seem to stay pretty constant as it makes progress:

total: 186863 blob: 93688 (6.6 per commit) commit: 14172 tree: 79003 (5.5 per commit)

Show 11 quoted lines

> For many commits, we generate as many (or more) _tree_ objects as we
> generate blobs.
>
> And tree obejcts from the same "supertree" really is something that I
> wouldn't mind packing some way, because they really tend to be very much
> related (since they refer to each other). Eg the commit and the top-level
> tree are almost always a pair, since you'd get a shared top-level tree
> only with two commits that have the exact same content (which definitely
> happens, don't get me wrong, but it we get some duplication for that case,
> we'd still be winning).
>

The packed item patch wouldn't duplicate info in this case. When it initially creates the packed buffer (before compression), it checks for an existing file with the same sha1 and returns if one is found. This is to preserve the optimizations for write_tree case where it frequently tries to create files that already exist.

-chris