On Sun, Jul 12, 2020 at 2:27 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:
> David Rowley <dgrowleyml@gmail.com> writes:
> > hmm yeah. It's unfortunate, but I'm not sure how I'd have implemented
> > it differently. The problem is made worse by the fact that we'll only
> > release the memory for the hash table during ExecEndHashJoin(). If the
> > planner had some ability to provide the executor with knowledge that
> > the node would never be rescanned, then the executor could release the
> > memory for the hash table after the join is complete.
>
> EXEC_FLAG_REWIND seems to fit the bill already?
FWIW I have a patch that does exactly that, which I was planning to
submit for CF2 along with some other patches that estimate and measure
peak executor memory usage.