Re: Performance issue in foreign-key-aware join estimation

On Fri, 21 Dec 2018 at 06:44, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:
> I was distressed to discover via perf that 69% of the runtime of this
> test now goes into match_eclasses_to_foreign_key_col(). That seems
> clearly unacceptable.

Agreed. That's pretty terrible.

I looked at this a bit and see that
match_eclasses_to_foreign_key_col() is missing any smarts to skip
equivalence classes that don't have ec_relids bits for both rels. With
that added the run-time is reduced pretty dramatically.

I've only tested with a debug build as of now, but I get:

Unpatched:

$ pgbench -n -T 60 -f query.sql postgres
latency average = 18411.604 ms

Patched:
latency average = 8748.177 ms

Going by my profiler this drops match_eclasses_to_foreign_key_col()
down to just 10% of total planner time for this query. The new
bms_is_member() call is pretty hot inside that function though.

--
David Rowley http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training & Services