Re: hash join hashtable size and work_mem

"Timothy J. Kordas" <tkordas@greenplum.com> writes:
> Am I missing something about the current implementation ?

If the planner has correctly predicted the number of rows, the table
loading should be about NTUP_PER_BUCKET in either regime.  Are you
sure you aren't just wishing that NTUP_PER_BUCKET were smaller?
I don't see that making the hashtable much larger than ntuples
is a good idea --- that just spreads out the live entries over more
cache lines, resulting in more cache thrashing.
        regards, tom lane