"Ross J. Reedstrom" <reedstrm@wallace.ece.rice.edu> writes:
> Ran across this paper, about an interesting approach, pulling in the indices
> to aid in selectivity estimates.
> http://db.cs.berkeley.edu/papers/CSD-98-1021.pdf
Looks pretty interesting, but also vastly more complex than I want to
tackle at the moment.
As of 6.5 the selectivity code is broken for everything except integers.
What I'm trying to do for this release cycle is to get it to operate as
well as can be expected given the existing design (in which available
statistics are not much more than min/max/mode values for each column;
strictly speaking the mode was not in the existing design, but given
that VACUUM was computing it anyway, we might as well use it).
A lot more *could* be done, as this paper suggests; but there are also
many other important tasks, and only so many hours in the day. I doubt
that building an entirely new selectivity estimation infrastructure is
worthwhile until we have cured some more problems elsewhere :-(
> while looking at the Mariposa work ( http://mariposa.cs.berkeley.edu)
> from the Sequoia2000 project. I've convinced my team to let me spend
> a couple days analyzing what it would take to fold the remote access
> features of Mariposa into the current PostgreSQL tree.
Let us know what you find out...
regards, tom lane
