Greg Stark <gsstark(at)mit(dot)edu> writes:
> The query actually runs slightly slower in CVS than with 7.3, though
> it's hard to compare because it seems to have done everything
> differently. Every hash join has become a merge join and every merge
> join has become a hash join :/
The estimated row counts for the bottom-level scans show some
differences, which probably account for the differing choice of plans.
Are these actually the same data with up-to-date ANALYZE stats in both
cases? I do not recall that we've made any adjustments to the basic
statistical estimation routines since 7.3 ...
regards, tom lane