On Wed, Mar 23, 2011 at 2:12 PM, Josh Berkus <josh@agliodbs.com> wrote:
> Folks,
>
>...
> It really seems like we should be able to detect an obvious high-risk
> situation like this one. Or maybe we're just being too optimistic about
> discarding subplans?
Why not letting the GEQO learn from past mistakes?
If somehow a post-mortem analysis of queries can be done and accounted
for, then these kinds of mistakes would be a one-time occurrence.
Ideas:
* you estimate cost IFF there's no past experience.
* if rowcount estimates miss by much, a correction cache could be
populated with extra (volatile - ie in shared memory) statistics
* or, if rowcount estimates miss by much, autoanalyze could be scheduled
* consider plan bailout: execute a tempting plan, if it takes too
long or its effective cost raises well above the expected cost, bail
to a safer plan
* account for worst-case performance when evaluating plans
