Alexey Nalbat <alexey(at)price(dot)ru> writes:
> Due to the updating algorithm the physical order of tuples in the
> table happens to be such that all equal keys are placed together,
> but not ordered globally.
Hmm... this is of course a weak spot of the correlation-based estimation
method. If you were doing a range query then the computed correlation
might have some bearing on the cost, but when probing for a single key
value, your table will behave much differently than the correlation model
can guess.
> Are there any better ideas?
None at the moment, but I'm open to suggestions. It seems like we might
need different stats for equality probes than range probes.
regards, tom lane