On Wed, Jul 31, 2019 at 11:07 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:
> What you'd want to do for something like the above, I think, is to
> have some kind of figure of merit or other special marking for paths
> that will have some possible special advantage in later planning
> steps. Then you can teach add_path that that's another dimension it
> should consider, in the same way that paths with different sort orders
> or parallizability attributes don't dominate each other.
Yeah, but I have to admit that this whole design makes me kinda
uncomfortable. Every time somebody comes up with a new figure of
merit, it increases not only the number of paths retained but also the
cost of comparing two paths to possibly reject one of them. A few
years ago, you came up with the (good) idea of rejecting some join
paths before actually creating the paths, and I wonder if we ought to
try to go further with that somehow. Or maybe, as Peter Geoghegan, has
been saying, we ought to think about planning top-down with
memoization instead of bottom up (yeah, I know that's a huge change).
It just feels like the whole idea of a list of paths ordered by cost
breaks down when there are so many ways that a not-cheapest path can
still be worth keeping. Not sure exactly what would be better, though.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company
