Peter Eisentraut wrote:
>
> mlw writes:
>
> > Adding huristics, such as weighting for index scans, is not making the planner
> > stupider. It is making it smarter and more flexable.
>
> If life was as simple as index or no index then this might make some
> sense. But in general the planner has a whole bunch of choices of join
> plans, sorts, scans, and the cost of an individual index scan is hidden
> down somewhere in the leaf nodes, so you can't simply say that plans of
> type X should be preferred when the cost estimates are close.
>
No doubt, no one is arguing that it is easy, but as I said in a branch of this
discussion, when the planner has multiple choices, and the cost ranges
overlapp, the relative numbers are not so meaningful that huristics would not
improve the algorithm.