Re: An unresolved performance problem.

Andrew Sullivan <andrew@libertyrms.info> writes:
> On Thu, May 08, 2003 at 10:48:52AM -0200, Achilleus Mantzios wrote:
>> That is, we have a marginal decrease of the total cost
>> for the index scan when random_page_cost = 1.9,
>> whereas the "real cost" in the means of total runtime
>> ranges from 218 msecs (seq scan) to 19 msecs (index scan).
>> (is it sane?)

> You're right that the problem is the poor estimate of the cost of
> that selection.

Are the table and index orders the same?  Oliver Elphick pointed out
awhile ago that we're doing a bad job of index order correlation
estimation for multi-column indexes --- the correlation is taken to
be much lower than it should be.  But if the correlation is near
zero anyway then this wouldn't explain Achilleus' problem...

            regards, tom lane