Michael Fuhr <mike@fuhr.org> writes:
> Limit (cost=0.00..25.79 rows=1 width=8) (actual time=631.964..631.964 rows=0 loops=1)
> -> Index Scan using foo_value_idx on foo (cost=0.00..2552.75 rows=99 width=8) (actual time=631.942..631.942
rows=0loops=1)
> Filter: (id = -1)
> Total runtime: 632.135 ms
> (4 rows)
> Maybe I don't understand something about what EXPLAIN is showing,
> but why does Limit have an estimated cost of 0.00..25.79 when the
> thing it's limiting has a cost of 0.00..2552.75?
This represents the planner assuming that the indexscan will only need
to be run 1/99th of the way to completion. That is, having estimated
that there were 99 matching rows to be found, it assumes those are
uniformly distributed in the index-by-value, and that the scan can stop
as soon as the first one is found.
Since in reality there aren't *any* matching rows, the index scan has to
go all the way to the end :-(. Even if there were matching rows, they
might be much further out in the index order than the
uniform-distribution hypothesis predicts, because the id and value
columns might have been correlated.
Basically, what you're looking at here is that the planner is thinking
it should go for a fast-start plan in a scenario where that bet loses.
It's still a good bet though. I'm not sure how to formulate the notion
that there's too much risk of a slow result in this scenario.
regards, tom lane