>
> "Mark Woodward" <pgsql@mohawksoft.com> writes:
>
>> The analyzer, at least the last time I checked, does not recognize these
>> relationships.
>
> The analyzer is imperfect but arguing from any particular imperfection is
> weak
> because someone will just come back and say we should work on that problem
> --
> though I note nobody's actually volunteering to do so whereas they appear
> to
> be for hints.
>
> I think the stronger argument is to say that there are some statistical
> properties that the analyzer _cannot_ be expected to figure out. Either
> because
>
> a) they're simply too complex to ever expect to be able to find
> automatically,
>
> b) too expensive to make it worthwhile in the general case, or
>
> c) because of some operational issue such as the data changing frequently
> enough that the analyzes that would be necessary to keep the statistics
> up
> to date would become excessively expensive or even be impossible to
> perform rapidly enough.
Well, from a purely data domain standpoint, it is impossible to charactize
the exact nature of a data set without enough information to recreate it.
Anything less must be designed for a fixed set of assumptions. There is no
way that every specific trend can be covered by a fixed number of
assumptions.
The argument that all we need is better statistics completely misses the
point. There will *always* be a number cases where the planner will not
work optimally. I would say that a "simpler" planner with better hints
will always be capable of creating a better query plan.