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

[Jakub Narebski <jnareb@gmail.com>]

On Wed, 6 Jul 2011, Jeff King wrote:
> On Wed, Jul 06, 2011 at 11:01:03AM -0400, Ted Ts'o wrote:
> 
> > Is it worth it to try to replicate this information across repositories?
> 
> Probably not. I suggested notes-cache just because the amount of code is
> very trivial.

Well, generation numbers are universal and would help everybody.  For
new commits with 'generation' header those would be always replicated,
for old commits with 'generation' notes / notes-cache the can be
replicated.
 
> One problem with notes storage is that it's not well optimized for tiny
> pieces of data like this (e.g., the generation number should fit in a
> 32-bit unsigned int, as its max is the size of the longest single path
> in the history graph). But notes are much more general; we will actually
> map each commit to a blob object containing the generation number, which
> is pretty wasteful.

Wasn't textconv-cache using commit-less notes?  The same can be done
for generation notes-cache.  Though it is still wasteful...  By the
way, would we be using text representation (like in 'generation'
commit header) or 32-bit integer binary representation in some
ordering, or variable-length integer (I think git uses them somewhere)?

Nb. I wonder if 32-bit unsigned int would always be enough, for example
Linux kernel + history.

> > Why not just simply have a cache file in the git directory which is
> > managed somewhat like gitk.cache; call it generation.cache?
> 
> Yeah, that would be fine. With a sorted list of binary sha1s and 32-bit
> generation numbers, you're talking about 24 bytes per commit. Or a 6
> megabyte cache for linux-2.6.
> 
> You'd probably want to be a little clever with updates. If I have
> calculated the generation number of every commit, and then do "git
> commit; git tag --contains HEAD", you probably don't want to rewrite the
> entire cache. You could probably journal a fixed number of entries in an
> unsorted file (or even in a parallel directory structure to loose
> objects), and then periodically write out the whole sorted list when the
> journal gets too big. Or choose a more clever data structure that can do
> in-place updates.

And that is the difference between gitk.cache (generated _once_ when starting
gitk, and regenerated on request), and idea of generation.cache

I think it would be simpler to use generation header + generation notes.
Or start with generation notes only.

-- 
Jakub Narebski
Poland