On Thu, Dec 7, 2017 at 6:01 AM, Amit Khandekar <amitdkhan...@gmail.com> wrote: > I have not managed to make the regression test cases execute the above > not-covered case. I could do that with my local test that I have, but > that test has lots of data, so it is slow, and not suitable for > regression. In select_parallel.sql, by the time a worker w1 gets into > the function choose_next_subplan_for_worker(), an earlier worker w2 > must have already wrapped around the pa_next_plan at place_2, so this > worker doesn't get a chance to hit place_1. It's a timing issue. I > will see if I can solve this.
Well, it seems to me that just running the test case with parallel_leader_participation = off might be good enough. It would hit a decent chunk of this logic, and that's better than what we have today. If we're determined to hit the wraparound case the way to do it is probably to make the first child of the append something that takes a long time to execute and the second one something quick. But, as you say, it's hard to do such things in the regression tests without making them slow, and I have observed Tom to dislike things that make the regression tests slow. -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company
