Re: [PATCH 0/3] batch blob diff generation

From: Jeff King <peff@peff.net>
Date: Dec 18, 2024, 11:39 UTC
Message-ID: <20241218113927.GB594795@coredump.intra.peff.net>
In-Reply-To: <xmqq5xnj7fre.fsf@gitster.g>

On Mon, Dec 16, 2024 at 08:29:41AM -0800, Junio C Hamano wrote:

Show 10 quoted lines

> Jeff King <peff@peff.net> writes:
> 
> > I recall that the performance of attr.tree is not great for _some_
> > commands (like pack-objects). So it's perhaps reasonable to use for
> > single commands like "git diff" but not to set in your on-disk config.
> > It's possible we'd want to warn people about that before advertising it
> > more widely? I dunno.
> 
> Or we disable the unusably-inefficient feature before doing so.
> Would attr.tree be much less efficient than GIT_ATTR_SOURCE?

Whether it's unusably inefficient depends on what you throw at it. IIRC, the performance difference for pack-objects on git.git was fairly negligible. The problem in linux.git is that besides being big, it has a deep(er) directory structure. So collecting all of the attributes for a file like drivers/gpu/drm/foo/bar.h needs to open all of those intermediate trees.

So I'd be inclined to leave it in place, in case somebody is actually happily using it.

GIT_ATTR_SOURCE suffers all of the same problems; it's just that you'd presumably only use it with a few select commands (as far as I know, pack-objects is the worst case because it's looking up one attribute on every single blob in all of history).

-Peff