on-the-fly, and thereby avoid blowing out memory when the planner has
underestimated the hash table size.  Hash join will now obey the
work_mem limit with some faithfulness.  Per my recent proposal
(hash aggregate part isn't done yet though).