"=?UTF-8?Q?Grzegorz_Ja=C5=9Bkiewicz?=" <gryzman@gmail.com> writes:
> so here are the plans, that's the real table run.
Hmm, well this rowcount estimate is way off:
> -> Hash Anti Join (cost=376.60..37791.22 rows=1
> width=8) (actual time=15.195..8216.448 rows=20000 loops=1)
The fact that it's getting a faster plan despite being completely wrong
about the rowcount means that the cost parameters are way off for your
situation. It looks like you are testing a case where the tables all
fit in memory. Do you expect that to be the reality for your production
use? If so, you might want to reduce random_page_cost to something
close to 1 to reflect it. If not, it'd be a good idea to test with more
realistically-sized tables before deciding what's "faster".
I'm not sure why the rowcount estimate is so far off, but the antijoin
code is all new and probably there's an estimation bug in there
somewhere. (You didn't get this plan, or anything at all like it,
from 8.1 ;-))
regards, tom lane