On Thu, 2009-05-28 at 15:12 -0400, Douglas Alan wrote:
> The most obvious solution would be an option to tell Postgres not to
> assume that the value is evenly distributed throughout the table, and
> to take account of the fact that the data in question might very well
> be clustered at the very end of the table.
There's no use adding a new statistic (user supplied or collected) to
PostgreSQL until we know that correlation isn't useful for that purpose.
Can you tell us the correlation that PostgreSQL has already collected (I
apologize if I missed this elsewhere in the thread)?
Hopefully, correlation is useful enough. With some planner tweaks
similar to the ones Tom mentioned, and a few more data points, maybe
we'll have a real solution.
Regards,
Jeff Davis