I was looking at inequality joins and I think I found an opportunity to
improve
how they are estimated.  The planner uses a fixed 1/3 (DEFAULT_INEQ_SEL) for
them, which is a sensible default when nothing is known, but these columns
often do have MCV lists and histograms, so I think we could use them.  For
a value v,
scalarineqsel() already computes the fraction of the left side satisfying
"x < v" for a constant; averaging that over the right side's distribution
turns
the join into an expectation, using statistics that are already there.  It
falls
back to 1/3 when a side's statistics cannot describe it, and seven tests
compare
the estimate with the actual row count.

I would appreciate feedback on the fallbacks and on the planning cost of
the extra lookups.

The code is here for now https://github.com/lrweck/postgres/pull/5

@Tomas Vondra <[email protected]>, if you think this approach has any ground,
I'd be more than
happy to continue pursuing it! I know I need to format it as a patch and
send it to
the hackers mailing list.

*Regards,*
*--*
*Luís R. Weck*

Reply via email to