Re: [HACKERS] Solution for LIMIT cost estimation

Don Baccus <dhogaza(at)pacifier(dot)com> writes:
>> The optimizer's job would be far simpler if no-brainer rules like
>> "indexscan is always better" worked.

> Yet the optimizer currently takes the no-brainer point-of-view that
> "indexscan is slow for tables much larger than the disk cache, therefore
> treat all tables as though they're much larger than the disk cache".

Ah, you haven't seen the (as-yet-uncommitted) optimizer changes I'm
working on ;-)

What I still lack is a believable approximation curve for cache hit
ratio vs. table-size-divided-by-cache-size. Anybody seen any papers
about that? I made up a plausible-shaped function but it'd be nice to
have something with some actual theory or measurement behind it...

(Of course the cache size is only a magic number in the absence of any
hard info about what the kernel is doing --- but at least it will
optimize big tables differently than small ones now.)

regards, tom lane