Re: Better estimates of index correlation

On 14.03.2011 16:09, Alvaro Herrera wrote:
> Excerpts from Joshua D. Drake's message of dom mar 13 23:20:01 -0300 2011:
>> On Sun, 2011-03-13 at 19:40 -0400, Tom Lane wrote:
>>
>>> I'm not planning to do anything about this idea right now, since I'm
>>> still hip-deep in collations, but I thought I'd throw it out to get
>>> it on the record.
>>>
>>> Comments?
>>
>> One question: Where is the overhead increase?
>
> During VACUUM, in the pass that processes indexes.
>
> I think Tom is sligthly confused though: AFAICT this must happen in
> btvacuumscan (which does the actual scan), not btvacuumcleanup (which
> may not do it, if btbulkdelete did it previously).  Which means it would
> be done for each pass over the index when vacuuming a relation, because
> I don't see any way for this function to determine whether this is the
> last pass we'll do over the index.
>
> It sure would be nice to be able to do it only during the last scan.

Can't we do it at ANALYZE? If the estimate is only based on intra-page 
comparisons anyway, a sample of random pages ought to be enough.

--   Heikki Linnakangas  EnterpriseDB   http://www.enterprisedb.com