What I'm thinking about in this thread is joins on TID, which we have only very weak support for today --- you'll basically always wind up with a mergejoin, which requires full-table scan and sort of its inputs. Still, that's better than a naive nestloop, and for years we've been figuring that that was good enough. Several people in the other thread that I cited felt that that isn't good enough. But if we think it's worth taking seriously, then IMO we need to add both parameterized scans (for nestloop-with-inner-fetch-by-tid) and hash join, because each of those can dominate depending on how many tuples you're joining.
That would certainly help if you are building a column store, or other new index types.
--
Simon Riggs http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services
