Re: generation numbers (was: [PATCH 0/4] Speed up git tag --contains)

From: Jakub Narebski <jnareb@gmail.com>
Date: Jul 6, 2011, 14:26 UTC
Message-ID: <m3mxgr4has.fsf_-_@localhost.localdomain>
In-Reply-To: <20110706070311.GA3790@sigill.intra.peff.net>

Jeff King <peff@peff.net> writes:

Show 20 quoted lines

> On Wed, Jul 06, 2011 at 01:56:23AM -0500, Jonathan Nieder wrote:
> > Jeff King wrote:
> > 
> > > The problem is that existing objects don't have this generation number.
> > > It's easy to calculate, though, and we could in theory use a notes-cache
> > > to store it externally. Obviously the complexity and performance aren't
> > > going to be as good as if it were just in the commit object, but we're
> > > sadly 6 years too late to make that decision.
> > 
> > I am still digesting the rest of what you wrote, but wouldn't this be
> > easy to do today?  One could just use a notes-cache while prototyping
> > and if it seems to work well, introduce new loose and packed object
> > formats that include a field for the cached generation number.
> 
> Yes, that's exactly how to do it. I'm just not sure "introduce new loose
> and packed object formats" is "easy to do". Though I'm not sure we need
> new formats. It is really just a new header in the commit object. And if
> we write the code carefully, we should be able to transparently use
> newly-generated objects with the field, and fall back to a notes-cache
> (with autogeneration) when it isn't there.

I understand that you would do autogeneration at least when you create a commit, and at least one of parents does not have generation number.

You can also autogenerate notes-cache when following commits, and encountering commit object without generation number.

Or make "git gc" autogenerate cache-notes for generation number,
perhaps with an option (i.e. probably not for "git gc --auto").

Show 12 quoted lines

> Existing git will ignore the new generation field. It does mean that old
> and new git will generate different sha1s for the exact same commit. I
> don't know how big a deal this is in practice. It matters a lot more for
> blobs and trees. But for commits, even if you are replaying a commit,
> you should be updating the commit timestamp, which is going to give a
> new sha1.
> 
> The other thing I worry about is performance. You are building a full
> notes tree and looking up every commit in the traversal. I don't know
> how bad that will be (though from my other back-of-the-envelope tests,
> it may not actually be that bad; notes were designed to be fast for
> exactly this case).

Well, one thing that it would test our notes infrastructure...

-- 
Jakub Narebski
Poland
ShadeHawk on #git