Re: What's cooking in git.git (Mar 2014, #01; Tue, 4)

From: Jeff King <peff@peff.net>
Date: Mar 14, 2014, 04:16 UTC
Message-ID: <20140314041618.GA32008@sigill.intra.peff.net>
In-Reply-To: <CACsJy8Do8nL1CcZ=jNz=_A2j8EWgcSAZBsqYZMF=SvTv0itRpg@mail.gmail.com>

On Wed, Mar 12, 2014 at 07:55:51PM +0700, Duy Nguyen wrote:

Show 10 quoted lines

> On Wed, Mar 5, 2014 at 7:10 AM, Junio C Hamano <gitster@pobox.com> wrote:
> > [Graduated to "master"]
> > * jk/pack-bitmap (2014-02-12) 26 commits
> >   (merged to 'next' on 2014-02-25 at 5f65d26)
> 
> And it's finally in! Shall we start thinking about the next on-disk
> format? It was put aside last time to focus on getting this series in.
> My concern is shallow support (surprise?) so that cloning from a
> 1-year-long shallow repo is not slower than a complete one. An
> extensible format is enough without going into details.

I'm open to suggestions. The most interesting followups for me are:

  1. Taking a look at Karsten's hashmap patches, which also graduated
     recently, to see if we can drop khash's macro nastiness without
     losing performance.

  2. Can we use bitmaps to increase our idea of which objects the other
     side has, and therefore get better delta reuse? I think the answer
     is yes from my limited experiments, but I want to do some more
     thorough ones[1].

  3. Cleaning up and submitting my patches to avoid writing deltas
     between "distinct" ref sets (e.g., between refs for two different
     forks in an alternates repo). This is somewhat orthogonal to
     bitmaps, but they touch overlapping areas, so I had held back until
     the dust settled on bitmaps.

  4. Packv4. :)

  5. Short of packv4, looking at whether we can encoding packv2 tree
     deltas in a way that lets us produce cheap tree diffs in a high
     percentage of instances. This may sound crazy, as packv4 is a
     better endgame, but there is going to be a long time where packv2
     is still the on-the-wire format, which means that servers will need
     to do v4-to-v2 translation, which I suspect may eliminate the v4
     benefits.

Shallow repos aren't really on my radar (nor am I really familiar with some of the intricacies), but I'm happy to discuss and give my ill-informed opinions. :)

[1] http://thread.gmane.org/gmane.comp.version-control.git/239647