On Fri, Apr 27, 2018 at 2:48 PM, Guilherme Pereira <guipe...@gmail.com> wrote: > Hi, > > Having a strange situation, where adding extra parallel workers > (max_parallel_workers_per_gather), the planner chooses a different plan,
I think I can see what is going on here. The planner chooses a different plan because the cost of that plan (cost=1001.10..31949141.56) is cheap as compared to the original plan (cost=31676816.72..32928717.16) > with nested loops, which makes the query twice as slow. It is slow because it needs to transmit many rows (rows=23845842) from workers to master backend. > Strangely with the > COUNT_DISTINCT implementation from Tomas Vondra > (https://github.com/tvondra/count_distinct) it scales nicely (almost > linearly) with the workers. > The main difference is that count_distinct is a parallel_safe aggregate function which can be allowed to push down to workers which in turn lead to very few rows (rows=9) being transferred from workers to master backend. I think ideally it should not have picked the plan which leads to transmitting so many rows, but the cost turns out to be lesser. -- With Regards, Amit Kapila. EnterpriseDB: http://www.enterprisedb.com
