Am Donnerstag, 6. März 2008 schrieb Tom Lane:
> What I propose doing about this is a small variant on Peter's original
> suggestion: compute the estimated selectivity for
> col = 'prefix'
> and clamp the result of prefix_selectivity to be at least that.
OK, first results with this patch are in: The selectivity estimations are
adjusted nicely, but the cost calculation doesn't change at all. Before:
Index Scan using foo_idx_3 on foo foo (cost=0.01..6.03 rows=1 width=8)
After:
Index Scan using foo_idx_3 on foo foo (cost=0.01..6.03 rows=627 width=8)
How is that possible?
Btw., the corresponding query plan for the LIKE 'constant' case is:
Index Scan using foo_idx_3 on foo foo (cost=0.00..2527.84 rows=627 width=8)
This is what we had hoped to get in the "after" case.