On Wed, Mar 22, 2017 at 12:55 AM, Robert Haas <robertmhaas@gmail.com> wrote: > On Tue, Mar 21, 2017 at 6:36 AM, Dilip Kumar <dilipbalaut@gmail.com> wrote: >> How about taking the decision of execute_once based on >> fcache->returnsSet instead of based on lazyEval? >> >> change >> + ExecutorRun(es->qd, ForwardScanDirection, count, !es->lazyEval); >> to >> + ExecutorRun(es->qd, ForwardScanDirection, count, !fcache->returnsSet); >> >> IMHO, Robert have the same thing in mind? > > Yeah, something like that. > Done in execute-once-v2.patch
So, let's see here. We are safe so long as we're sure that, when postquel_getnext() returns, postquel_end() will be called next without iterating the loop in fmgr_sql(). That will definitely be true if fcache->returnsSet is true. It will also be true if postquel_getnext returns true, which will be true whenever count == 0, which will be true whenever es->lazyEval is false.
So couldn't we actually make this test !fcache->returnsSet || !es->lazyEval? That would let us allow parallel execution for all non-set-returning functions, and also for set-returning functions that end up with es->lazyEval set to false.