wieck@debis.com (Jan Wieck) writes:
> Up to now (v6.5) this kind of nested aggregates isn't
> supported. Not directly over SQL, nor by views. To be sure
> anything is fine, your views (and however you select from
> them) should be expressable with a regular SELECT too. In
> fact the rewrite system has to try to build such a query for
> it - so if you can't how should the rewriter can?
Still, it ought to either do it or produce a useful error message.
6.4's error message doesn't qualify as useful in my book. But 6.5's
behavior is far worse: it accepts the query and cheerfully generates
a wrong result! That's definitely a bug.
Looking ahead to the larger problem, I believe that the executor is
perfectly capable of handling nested aggregate plans --- the trick is
to get the planner to produce one. Maybe we need an extension to the
parsetree language? It doesn't seem like this ought to be hard to
support, it's just that there's no parsetree configuration that
represents what we want done. Or, maybe we should rethink the division
of labor between the rewriter and planner --- if the rewriter could
output a partially-converted plan tree, instead of a parse tree, then
it could do as it pleased, but still leave the messy details of lowlevel
plan optimization to the planner.
regards, tom lane