On Fri, 26 Jan 2024 at 20:42, Richard Guo <guofengli...@gmail.com> wrote: > > In 5543677ec9 we introduced an optimization that uses Limit instead of > Unique to implement DISTINCT when all the DISTINCT pathkeys have been > marked as redundant. I happened to notice that this optimization was > not applied to partial DISTINCT, which I think should be.
It seems very likely that the parallel plan would only be chosen if the planner estimated there'd just be 1 row before the distinct. Otherwise, the non-partial path's LIMIT would come out so cheap that it would be unlikely that the parallel plan would be picked. I think your test case only chooses the parallel plan because you're doing FROM tenk1 WHERE four=4. And that column only contains values 0..3. However, having said that. Parallel plans are often picked when there is some highly selective qual as parallel_tuple_cost has to be applied to fewer tuples for such plans, so probably this is worth doing. David