On Monday, March 17, 2014, Tom Lane <t...@sss.pgh.pa.us> wrote: > Claudio Freire <klaussfre...@gmail.com <javascript:;>> writes: > > On Mon, Mar 17, 2014 at 7:01 PM, Jim Nasby <j...@nasby.net <javascript:;>> > wrote: > >> Even better would be if the planner could estimate how bad a plan will > >> become if we made assumptions that turn out to be wrong. > > > That's precisely what risk estimation was about. > > Yeah. I would like to see the planner's cost estimates extended to > include some sort of uncertainty estimate, whereupon risk-averse people > could ask it to prefer low-uncertainty plans over high-uncertainty ones > (the plans we typically choose for ORDER BY ... LIMIT queries being great > examples of the latter). But it's a long way from wishing that to making > it so. Right now it's not even clear (to me anyway) how we'd measure or > model such uncertainty. >
Most of the cases where I've run into horrible estimates, it seemed like the same level of knowledge/reasoning that could allow us to know it was risky, would allow us to just do a better job in the first place. The exception I can think of is in an antijoin between two huge rels. It is like subtracting two large measurements to get a much smaller result. We should know the uncertainty will be large. Cheers, Jeff