Re: Preferring shallower deltas on repack

From: Nicolas Pitre <nico@cam.org>
Date: Jul 9, 2007, 19:13 UTC
Message-ID: <alpine.LFD.0.999.0707091503000.26459@xanadu.home>
In-Reply-To: <20070709185353.GL4087@lavos.net>

On Mon, 9 Jul 2007, Brian Downing wrote:

Show 13 quoted lines

> On Mon, Jul 09, 2007 at 09:39:44AM -0700, Junio C Hamano wrote:
> > > OK here it is.  And results on the GIT repo and another patalogical test 
> > > repo I keep around are actually really nice!  Not only the pack itself 
> > > is a bit smaller, but the delta depth distribution as shown by 
> > > git-verify-pack -v is much nicer with the bulk of deltas in the low 
> > > depth end of the spectrum and no more peak at the max depth level.
> > 
> > Looks obviously correct.  Brian, it would be very interesting to
> > see what Nico's patch does to your dataset.
> 
> Nico's patch makes the overall statistics look better, but the
> version.lisp-expr file still goes 593 levels deep, as opposed to about
> 65 with my patch.  (That's better than 980 with stock Git, though.)

Tuning for such an extreme without impacting normal cases is rather hard.

My patch was meant to be used on top of yours.  Is that what you tested?

Also I'd suggest you do not use a max depth of 1000. It is simply insane and might possibly make the existing logic less effective. Even for runtime pack access you want it to be reasonably short, say 100 maximum, or even the current default of 50. Any improvements you might come with (like automatic depth determined on replay cost) should be compared with that default which is known to work well already.

Nicolas