Andrew Gierth <andrew@tao11.riddles.org.uk> writes:
> "Alexander" == Alexander Kuzmenkov <a.kuzmenkov@postgrespro.ru> writes:
> Alexander> Structurally, the patch consists of two major parts: a
> Alexander> specialized executor node
> Why?
> It strikes me that the significant fact here is not that we're doing
> count(*), but that we don't need any columns from the bitmap heap scan
> result. Rather than creating a whole new node, can't the existing
> bitmap heapscan be taught to skip fetching the actual table page in
> cases where it's all-visible, not lossy, and no columns are needed?
+1 ... while I hadn't actually looked at the code, it seemed to me that
anything like the optimization-as-described would be impossibly klugy
from the planner's standpoint. Your formulation sounds lots nicer.
Detecting that no columns are needed in the executor might be a bit tricky
because of the planner's habit of inserting a "physical tlist" to avoid a
projection step. (See also nearby discussion about custom scan planning.)
But we could fix that. I think one rule that would make sense is to
just disable the physical-tlist substitution if the relation's targetlist
is empty --- it wouldn't be buying much in such a case anyhow. Then the
runtime tlist for the scan node would also be empty, and away you go.
regards, tom lane