On Wed, May 12, 2021 at 6:30 PM Bharath Rupireddy <bharath.rupireddyforpostg...@gmail.com> wrote: > > On Wed, May 12, 2021 at 9:48 AM Amit Kapila <amit.kapil...@gmail.com> wrote: > > > > On Tue, May 11, 2021 at 6:31 PM Justin Pryzby <pry...@telsasoft.com> wrote: > > > > > > On Tue, May 11, 2021 at 05:37:50PM +0530, Bharath Rupireddy wrote: > > > > 3) Should the Assert(nindexes > 0); in begin_parallel_vacuum just be > > > > Assert(nindexes > 1); as this function is entered only when indexes > > > > are > 1? > > > > > > I think you're right, at least with the current implementation that > > > parallelization is done across indexes. Same in parallel_vacuum_main. > > > > > > > Yeah, as code stands both of you are right. However, it can be helpful > > to test parallelism even with one index say if we implement something > > like force_parallel_mode = regress or parallel_leader_participation = > > off. > > I see that currently we don't have it yet. Is it worth implementing > them? Something like 1) when force_parallel_mode = regress, spawn one > parallel worker, send the relation information to it, so that it > performs vacuuming both the relation and it's indexes. 2) > parallel_leader_participation = off, spawn workers as specified, but > don't let the leader to vacuum index, so that any worker can pick it > up. I'm not sure of the complexity though. >
We had some patch on the above lines in the original development thread which has been used for testing during development but not sure how much useful it is now. However, I am fine if others think something like that is useful. -- With Regards, Amit Kapila.
