Alvaro Herrera <alvherre@2ndquadrant.com> writes:
> Tom Lane wrote:
>> Yeah, fixed. I had assumed that the existing coding in create_gather_plan
>> was OK, because it looked like it was right for a non-projecting node.
>> But actually Gather can project (why though?), so it's not right.
> This looks related:
> https://www.postgresql.org/message-id/CA%2BTgmoai9Ruhwk61110rY4cNAzoO6npsFEOaEKjM7Zz8i3evHg%40mail.gmail.com
Ah, thanks for remembering that thread; I'd forgotten it.
Gather is a bit weird, because although it can project (and needs to,
per the example of needing to compute a non-parallel-safe function),
you would rather push down as much work as possible to the child node;
and doing so is semantically OK for parallel-safe functions. (Pushing
functions down past a Sort node, for a counterexample, is not so OK
if you are concerned about function evaluation order, or even number
of executions.)
In the current code structure it would perhaps be reasonable to teach
apply_projection_to_path about that --- although this would require
logic to separate parallel-safe and non-parallel-safe subexpressions,
which doesn't quite seem like something apply_projection_to_path
should be doing.
This seems rather closely related to the discussion around Konstantin
Knizhnik's patch to delay function evaluations to after the ORDER BY
sort when possible/useful. What I think that patch should be doing is
for grouping_planner (or subroutines thereof) to classify tlist items
as volatile or expensive or not and generate pre-sort and post-sort
targetlists appropriately. That's okay for a top-level feature like
ORDER BY, but it doesn't work for Gather which can appear much further
down in the plan. So maybe apply_projection_to_path is the only
feasible place.
regards, tom lane
