"William N. Zanatta" <william@veritel.com.br> writes:
> OK, as I don't know what is much and what isn't for the stats target,
> I tried to increase it firstly to 15 (10 is default according to
> documentation) and then to 20. Results follow:
Try like 50 or 100. Also, does the EXPLAIN result change with the
higher correlation figures? I'd expect estimated cost of indexscan
to drop with higher correlation.
If you dig into the pghackers archives you will find some discussions
about changing the equation used for adjusting indexscan based on
correlation. I suspect the real solution to your problem will be to
alter the equation. But I'm curious about whether we have reliable
stats, first...
regards, tom lane