Re: Delete query takes exorbitant amount of time

On Tue, 2005-03-29 at 09:40 -0500, Tom Lane wrote:
> Simon Riggs <simon@2ndquadrant.com> writes:
> > ...but, I see no way for OidFunctionCall8 to ever return an answer of
> > "always just 1 row, no matter how big the relation"...so tuples_fetched
> > is always proportional to the size of the relation. Are unique indexes
> > treated just as very-low-selectivity indexes?
>
> Yeah.  It is not the job of amcostestimate to estimate the number of
> rows, only the index access cost.  (IIRC there is someplace in the
> planner that explicitly considers unique indexes as a part of developing
> selectivity estimates ... but it's not that part.)

Well, I mention this because costsize.c:cost_index *does* calculate the
number of rows returned. If unique indexes are handled elsewhere then
this would not cause problems for them...but for LIMIT queries..?

cost_index gets the selectivity then multiplies that by number of tuples
in the relation to calc tuples_fetched, so it can use that in the
Mackert & Lohman formula. There's no consideration of the query limits.

That implies to me that LIMIT queries are not considered correctly in
the M&L formula and thus we are more likely to calculate a too-high cost
for using an index in those circumstances....and thus more likely to
SeqScan for medium sized relations?

Best Regards, Simon Riggs