FWIW, I've repeated the TPC-DS tests on a much larger data set (50GB)
today, and I see that
(a) 3f59be836c555fa679bbe0ec76de50a8b5cb23e0 (ANTI/SEMI join costing) changes nothing - there are some small cost
changes,but only in plans involving semi/anti-joins (which is expected). Nevertheless, all the plans remain the
same.
(b) 3b0f77601b9f9f3a2e36a813e4cd32c00e0864d6 (add_path fixes)
This changes join order in one of the queries, with lots of nested loops (this is the join order change we've
seenin this thread). Anyway, this is mostly expected consequence of the add_path changes.
So both changes seem fine.
--
Tomas Vondra http://www.2ndQuadrant.com
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services