I wrote:
> I instrumented the code in setrefs.c, and found that during the
> core regression tests this patch estimates correctly in 2103
> places while guessing wrongly in 54, so that seems like a pretty
> good step forward.
On second thought, that's not a terribly helpful summary. Breaking
things down to the next level, there were
1088 places where we correctly guessed a subquery isn't trivial
(so no change from current behavior, which is correct)
1015 places where we correctly guessed a subquery is trivial
(hence, improving the cost estimate from before)
40 places where we incorrectly guessed a subquery isn't trivial
(so no change from current behavior, although that's wrong)
14 places where we incorrectly guessed a subquery is trivial
(hence, incorrectly charging zero for the SubqueryScan)
1015 improvements to 14 disimprovements isn't a bad score. I'm
a bit surprised there are that many removable SubqueryScans TBH;
maybe that's an artifact of all the "SELECT *" queries.
regards, tom lane