Re: Horrible re-packing?

From: Nicolas Pitre <nico@cam.org>
Date: Jun 5, 2006, 23:13 UTC
Message-ID: <Pine.LNX.4.64.0606051900170.24152@localhost.localdomain>
In-Reply-To: <Pine.LNX.4.64.0606051432270.5498@g5.osdl.org>

On Mon, 5 Jun 2006, Linus Torvalds wrote:

Show 14 quoted lines

> 
> 
> On Mon, 5 Jun 2006, Nicolas Pitre wrote:
> > 
> > In other words, the pack shrunk to less than half the size of the 
> > previous one !
> 
> Ok, that's a bit more extreme than expected.
> 
> It's obviously great news, and says that the approach of sorting by 
> "reversed name" is a great heuristic, but at the same time it makes me 
> worry a bit that this thing that is supposed to be a heuristic ends up 
> being _so_ important from a pack size standpoint. I was happier when it 
> was more about saving a couple of percent.

Well... this is the repository that exhibited a repack regression a while ago, going from something like ~40MB to ~160MB when Junio initially added the directory in the name hash. No other popular repositories had that problem.

Which is why I said this repo is particularly sensitive to heuristic changes. So I wouldn't worry too much about your proposed patch making it too great in this case. It certainly didn't cause any (significant) regression overall which is what matters.

We already have surprizing results when combining two heuristics together although if used separately they do worse. So trying to have fallback/incremental heuristics is going to make things simply too complicated for when it breaks. Better experiment with new ideas and adopt them when they do a better job universally.

... which your proposed hashing change does.

Nicolas