On Thu, 2006-11-09 at 04:09 -0500, Andrew Sullivan wrote:
> On Mon, Nov 06, 2006 at 09:50:53PM +0000, Simon Riggs wrote:
> > - There are specific issues with the optimizer's ability to understand
> > dead row numbers, which can in some cases lead to SeqScan plans that are
> > inappropriate when tables grow because of updates. This is a red-herring
> > that can lead to people thinking the situation is worse than it is; that
> > needs fixing, but the core issues mentioned above remain.
>
> I don't disagree with much of what you say, but I'm slightly
> concerned about the wave-away answer that you give here. In my
> experience on high-update tables -- especially ones with the ones
> with few rows, but lots of open transactions over the lifetime of the
> row -- accurate understanding of dead rows would be a _dramatic_
> improvement (perhaps at least as significant as the improvement being
> discussed).
Understood. I just wanted to make sure people understand that the
underlying problem would still be there even if we fix that.
> That said, I'm not opposed to the line you're taking. I just don't
> want this problem to sink forever, because it's a big problem.
So, yeh, we should still fix that. The current prototype has a different
cost model for SeqScans as a result. Summary arriving anytime now.
-- Simon Riggs EnterpriseDB http://www.enterprisedb.com
