Re: [PATCH v3 0/8] Fix GIT_CEILING_DIRECTORIES that contain symlinks

From: David Aguilar <davvid@gmail.com>
Date: Oct 29, 2012, 00:15 UTC
Message-ID: <CAJDDKr4ki+NjSeuZpAU6bM=YAQ_3mdHCtawstdCqe9Ewvp=arQ@mail.gmail.com>
In-Reply-To: <7v7gqkgvxe.fsf@alter.siamese.dyndns.org>

On Sat, Oct 20, 2012 at 11:51 PM, Junio C Hamano <gitster@pobox.com> wrote:

Show 23 quoted lines

> Michael Haggerty <mhagger@alum.mit.edu> writes:
>
>> This patch series has the side effect that all of the directories
>> listed in GIT_CEILING_DIRECTORIES are accessed *unconditionally* to
>> resolve any symlinks that are present in their paths.  It is
>> admittedly odd that a feature intended to avoid accessing expensive
>> directories would now *intentionally* access directories near the
>> expensive ones.  In the above scenario this shouldn't be a problem,
>> because /home would be the directory listed in
>> GIT_CEILING_DIRECTORIES, and accessing /home itself shouldn't be
>> expensive.
>
> Interesting observation.  In the last sentence, "accessing /home"
> does not exactly mean accessing /home, but accessing / to learn
> about "home" in it, no?
>
>> But there might be other scenarios for which this patch
>> series causes a performance regression.
>
> Yeah, after merging this to 'next', we should ask people who care
> about CEILING to test it sufficiently.
>
> Thanks for rerolling.

GIT_CEILING_DIRECTORIES was always about trying to avoid hitting them at all; they can be (busy) NFS volumes there.

Here's the description from the 1.6.0 release notes:

* A new environment variable GIT_CEILING_DIRECTORIES can be used to stop
  the discovery process of the toplevel of working tree; this may be useful
  when you are working in a slow network disk and are outside any working tree,
  as bash-completion and "git help" may still need to run in these places.

In 8030e44215fe8f34edd57d711a35f2f0f97a0423 Lars added GIT_ONE_FILESYSTEM to fix a related issue. Do you guys have GIT_CEILING_DIRECTORIES set too?

We use GIT_CEILING_DIRECTORIES and I'm pretty sure we don't want every git command hitting them, so this would be a regression when seen from the POV of our current usage of this variable, which would be a bummer.

Is there another way to accomplish this without the performance hit? Maybe something that can be solved with configuration?

I'd be happy to lend a hand if you guys have some ideas on how we can help keep it fast. Thoughts?

Original patches for those just joining us: http://thread.gmane.org/gmane.comp.version-control.git/208102

-- 
David