Re: [PATCH] check_everything_connected: assume alternate ref tips are valid

From: Jeff King <peff@peff.net>
Date: Jun 29, 2019, 07:55 UTC
Message-ID: <20190629075536.GB5080@sigill.intra.peff.net>
In-Reply-To: <xmqqpnmx8ysf.fsf@gitster-ct.c.googlers.com>

On Fri, Jun 28, 2019 at 09:22:56AM -0700, Junio C Hamano wrote:

Show 10 quoted lines

> >  	argv_array_push(&rev_list.args, "--quiet");
> > +	argv_array_push(&rev_list.args, "--alternate-refs");
> >  	if (opt->progress)
> >  		argv_array_pushf(&rev_list.args, "--progress=%s",
> >  				 _("Checking connectivity"));
> 
> Quite honestly, I am very surprised that we did not do this.  The
> idea of alternate object store, as well as reducing transfer cost by
> advertising their tips as '.have' phony refs, is almost as old as
> the pack protocol itself.

Yeah, as you note we are already telling the other side of the push "hey, we already have these objects". So we are almost always just walking over our own local objects in the connectivity check, which is silly.

I only did "clone --reference" in the perf test because it was the simplest, but a push to a server with alternates should be similarly improved. E.g., doing this in a clone of linux.git:

  git init --bare dst.git
  echo '../../.git/objects' >dst.git/objects/info/alternates
  time git push dst.git HEAD

goes from 40+ seconds to 100ms or so. Again, obviously that's the best case, but it should also improve the normal case of somebody pulling down "torvalds/linux.git" and pushing it back up to their own "peff/linux.git", too.

I don't have real-world numbers yet from GitHub, because we're not actually advertising .haves on push right now. All of the Git pieces are now in places to do so, but we still have to make some tweaks at our replication layer. But soon. :)

-Peff