On 10/25/24 02:43, Ba Jinsheng wrote:
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?
I also discovered your case. Using AQO and settling the correct
cardinalities in each node, I found that the plan doesn't change at all.
So, I wonder if you could analyse the path-choosing logic, determine the
costs of competing paths, and explain why NestLoop wasn't chosen.
Maybe there is kind of early selectivity estimation error or something
even more deep: specific tuples distribution across blocks of the heap
table.
--
regards, Andrei Lepikhov
