Ron Johnson <ron(dot)l(dot)johnson(at)cox(dot)net> writes:
> How about (if you don't already do it) ranked (or approximately
> ranked) b-tree indexes, where each node also stores the (approximate)
> count of tuple pointers under it?
> This way, the planner would know whether or how skewed a tree is,
> and (approximately) how many tuples a given WHERE predicate resolves
> to.
Why is that better than our existing implementation of column statistics?
regards, tom lane