On Thu, 10 Dec 2009 15:00:42 +0800, Kan-Ru Chen <kanru@kanru.info> wrote:
> The termlist is already sorted, so this is the patch trying to minimize
> the modification of database as suggested in the comment and Carl's
> TODO file.

Fantastic, Kan-Ru!

> My poor profiling shows not much, but some improvement.

Now you're just understating for sake of the pun. A 5-6x performance
improvement looks great. And I see that as well in my testing:

Before:

	$ time notmuch tag +foo tag:sent
	real 3m18.067s
	$ time notmuch tag -foo tag:sent
	real 3m12.940s

After:

	$ time notmuch tag +foo tag:sent
        real 0m31.497s
	$ time notmuch tag -foo tag:sent
	real 0m28.722s

I didn't flush the OS caches between runs, but a subsequent run of the
"before" code still performed similarly slow:

	$ time notmuch tag +foo tag:sent
	real 3m7.172s

And if I *had* used cold caches for every run the benefit of the patch
would have looked even better.

Anyway, we should get this upstream to the Xapian folks straight
away. I expect they'll want to see a patch to the chert backend as well
as the flint backend, (but fortunately the relevant code looks very
similar if not identical).

Thanks again,

-Carl