Ba Jinsheng <bajinsheng@xxxxxxxxx> writes:
> I am not proposing a fixing patch, as the patch is incorrect. Instead, I just want to show disabling the simplify_function() function brings performance benefit, and it seems unexpected. I am wondering whether we can optimize simplify_function() to make the performance better for this workload?
The improvement here seems purely accidental. Basically, you have
broken the rowcount estimation for the "orders" scan: instead of
> -> Parallel Seq Scan on orders (cost=0.00..35511.00 rows=23858 width=8) (actual time=0.033..62.907 rows=19046 loops=3)
> Filter: ((o_orderdate >= '1993-08-01'::date) AND (o_orderdate < '1993-11-01 00:00:00'::timestamp without time zone))
> Rows Removed by Filter: 480954
we get
> -> Parallel Seq Scan on orders (cost=0.00..37062.50 rows=3125 width=8) (actual time=0.029..63.908 rows=19046 loops=3)
> Filter: ((o_orderdate >= '1993-08-01'::date) AND (o_orderdate < ('1993-08-01'::date + '3 mons'::interval month)))
> Rows Removed by Filter: 480954
The rows estimate is a good deal further from reality, as a
consequence of the fact that scalarltsel is now delivering a default
guess instead of a real estimate based on a constant comparison value.
Somehow, that's nonetheless leading to a better plan choice. It could
be that the underlying problem is a poor cost model for parallel hash
join. Or it could be as simple as random_page_cost not being tuned
for your environment. It looks like the better plan involves a
nestloop with inner indexscan on lineitem, which is something whose
estimated cost depends enormously on random_page_cost. You've given
us exactly zero detail about your test conditions, so it's hard to say
more than that.
regards, tom lane
