"Marc G. Fournier" <scrappy@hub.org> writes:
> [ bogus optimizer choices in 7.2 ]
Well, I guess the good news is we seem to be past the old bugaboo of bad
statistics: the estimated row counts are all in the right ballpark. Now
we get to have fun with the cost models :-).
It looks to me like there are a couple of problems here. One is that
the default value of effective_cache_size is way too small --- it's set
at 1000, which is probably silly when you have NBuffers set to 32768.
(In hindsight maybe we should have expressed it as a multiple of
NBuffers rather than an absolute size.) You could tweak that with a
postgresql.conf change, but I'm not sure that that alone will help much.
The more difficult issue is that nestloops with inner indexscan are
being seriously misestimated. We're computing the cost as though each
iteration of the inner scan were completely independent and being done
from a standing start --- which is wrong, because in practice scans
after the first will tend to find buffer cache hits for pages already
read in by prior scans. You can bet, for example, that the btree
metapage and root page aren't going to need to be re-read on each
iteration.
I am thinking that the right way to do this is to cost the entire inner
indexscan (all iterations put together) as if it were a single
indexscan, at least for the purposes of applying the Mackert & Lohman
formula embedded in cost_index. That would give us a more realistic
result for the total cost of the main-table accesses driven by the
index. Not sure how to adjust the cost estimate for reading the index,
but clearly we need to make some adjustment for repeated hits on the
upper index pages.
This is probably a bigger change than we can hope to make in 7.2.* ...
BTW, what do you get if you EXPLAIN ANALYZE that orient/clubs join
with seqscan enabled and hashjoin disabled? If it's a mergejoin,
how about if you also disable mergejoin? It seems to me that a seqscan
on clubs would be a much better way to do the nestloop join than an
indexscan --- but it's being forced into an indexscan because you
disabled seqscan.
regards, tom lane
