2010/12/27 Tomas Vondra <tv@fuzzy.cz>:
> But even though these disadvantages, there really is no other
> way to enhance the estimates. I don't think this should be a
> default behavior - just as in case of cross-column stats this should
> be optional when the current estimator does not work well.
This is going to be a lot of work to implement, so before you do it,
we should try to reach a consensus that (a) it's part of an overall
strategy that the community generally supports and (b) we have
consensus on the design for this part.
With respect to (a), I have to admit I've found the discussion on
cross-column stats to be quite difficult to follow. I'd like to see a
very simple description of exactly what information we're going to
store, under what circumstances we'll store it, and how we'll use it
to compute selectivity estimates.
With respect to (b), I think I'd need to see a much more detailed
design for how you intend to make this work. Off the top of my head
there seems to be some pretty serious feasibility problems.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company