On Wed, 2007-12-05 at 15:13 -0500, Chris Browne wrote:
> I have the theory (thus far not borne out by any numbers) that it
> might be a useful approach to try to go through the DB schema and use
> what information is there to try to come up with better numbers on a
> per-column basis.
Yeh, agreed.
The difficulty is making this work for generic datatypes.
> - Datestamps tend to imply temporal dispersion, ergo "somewhat fewer
> bins." Similar for floats.
Hmmm, not sure about that one. Some date/time columns can change very
quickly over time, so the stats are frequently out of date.
> Then could come a "second order" perspective, where data would
> actually get sampled from pg_statistics.
>
> - If we look at the number of distinct histogram bins used, for a
> particular column, and find that there are some not used, we might
> drop bins.
The histograms are height balanced, so they are always all used.
--
Simon Riggs
2ndQuadrant http://www.2ndQuadrant.com