Re: Window Functions with identical PARTITION BY and ORDER BY clauses evaluated separately

Christopher Inokuchi <cinokuchi(at)gmail(dot)com> writes:
> Was it really not intentional that the docs explicitly name PARTITION BY
> and ORDER BY rather than the entire window_definition? If I understand
> correctly, only those two clauses control which records are hit and in what
> order.

Yeah, it's intentional, and in fact required by the SQL standard.
However, you're misinterpreting what the guarantee is. The spec
requirement is that window functions sharing PARTITION BY and
ORDER BY all be evaluated on the same concrete ordering of the
data, ie there can't be any re-sorting between them. And that's
what we implement. We do use a separate WindowAgg node for
each distinguishable window specification, but you'll notice
there is not a Sort step between them unless the query involves
entirely-incompatible PARTITION/ORDER BY specs.

Perhaps the wording in section 7.2.5 could be improved; I agree
that "evaluated in one pass" is capable of being read in more
than one way, and it's not clear that it's referring to sorts.
Do you have any suggestions for clearer wording?

It's possible that we could restructure things so that window
functions having distinct frame clauses were nonetheless done
in one WindowAgg node. But it would complicate the code and
it's far from obvious to me that it'd buy much in speed.
Optimizing the sort steps is where most of the potential win lies.

regards, tom lane