Re: An unresolved performance problem.

Achilleus Mantzios kirjutas K, 07.05.2003 kell 19:33:
> Hi, few days ago, i posted some really wierd (at least to me)
> situation (maybe a potentian bug) to the performance and bugs list
> and to some core hacker(s) privately as well,
> and i got no response.
> Moreover i asked for some feedback
> in order to understand/fix the problem myself,
> and again received no response.
>
> What i asked was pretty simple:
> "1. Is it possible that the absense of statistics make the planer produce
> better plans than in the case of statistcs generated with vacuum
> analyze/analyze?

Yes, the planner is not perfect, the statistics are just statistics
(based on a random sample), etc..

This question comes up at least once a month on either [PERFORM] or
[HACKERS], search the mailing lists to get more thorough
discussion/explanation.

> 2. If No, i found a bug,

Rather a feature ;-p

> 3. If yes then under what conditions??

if

1) ANALYZE produced skewed data which was worse than default.

or.

2) some costs are way off for your system (try changing them in
postgresql.conf)

> 4. If no person knows the answer or no hacker wants to dig into the
> problem then is there a direction i must follow to understand/fix whats
> going on myself??""

You can sturt by enabling/disabling various scan methods

psqldb# set enable_seqscan to off;
SET

and see what happens, then adjust the weights in postgresql.conf or use
some combination of SETs around critical queries to force the plan you
like.

------------
Hannu