Robert Haas <robertmhaas@gmail.com> writes:
> On Thu, Jun 30, 2016 at 5:54 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
>> And the point of that is what, exactly? If the only change is that
>> "some restrictions get enforced", I am not clear on why we need such
>> a test mode in cases where the planner is afraid to put a top Gather on
>> the plan. In particular, given the coding as you now have it, it seems
>> like the only case where there's any difference is where we set
>> glob->parallelModeOK but nonetheless end up with a not-parallel-safe
>> topmost path (that doesn't have a Gather within it). It's not clear
>> to me why having the executor switch into parallel mode makes sense at
>> all with such a plan.
> Suppose you create a PL/pgsql function that does an UPDATE and mark it
> PARALLEL RESTRICTED. You wonder whether you've marked it correctly.
> You can set force_parallel_mode=on and SELECT myfunc(). The
> subsequent ERROR tells you that you've mismarked it.
Right, but that statement is still true with the logic I'm imagining.
I would also argue that the existing text in config.sgml explaining
what this parameter does corresponds much more nearly to what I'm
suggesting than to what you say the semantics are.
>> What I'm not happy about is that as you've got things constituted,
>> the GetForeignUpperPaths hook is broken so far as injecting parallel paths
>> is concerned, because the consider_parallel flags for the upperrels
>> aren't set yet when it gets called. If we keep consider_parallel with
>> its present usage, we're going to have to refactor things to fix that.
> I see. It seems to me, and I may be failing to understand something,
> that the placement of create_upper_paths_hook is substantially better
> than the placement of GetForeignUpperPaths. If the latter were moved
> to where the former now is, then consider_parallel would be set
> sufficiently early and everything would be fine.
Yeah, I came to more or less the same conclusion last night. Will see
to it after you commit this.
regards, tom lane