On Mon, Mar 14, 2016 at 9:18 AM, Amit Kapila <amit.kapila16@gmail.com> wrote:
> On Fri, Feb 26, 2016 at 4:37 PM, Robert Haas <robertmhaas@gmail.com> wrote:
>> And, I'm going to revert this part. If you'd run the regression tests
>> under force_parallel_mode=regress, max_parallel_degree>0, you would
>> have noticed that this part breaks it, because of CREATE TABLE ... AS
>> EXECUTE.
>>
>
> I have looked into this issue and found that the reason for the failure is
> that in force_parallel_mode=regress, we enable parallel mode restrictions
> even if the parallel plan is not choosen as part of below code in
> standard_planner()
>
> if (force_parallel_mode == FORCE_PARALLEL_OFF || !glob->parallelModeOK)
>
> {
>
> glob->parallelModeNeeded = false;
>
> glob->wholePlanParallelSafe = false; /* either false or don't care */
>
> }
>
> else
>
> {
>
> glob->parallelModeNeeded = true;
>
> glob->wholePlanParallelSafe =
>
> !has_parallel_hazard((Node *) parse, false);
>
> }
>
>
> The failure cases fall into that category, basically wholePlanParallelSafe
> will be false, but parallelModeNeeded will be true which will enable
> parallel mode restrictions even though the plan won't contain Gather node.
> I think if we want to operate in above way for testing purpose, then we need
> to force during execution that statements for non read-only operations
> should not enter into parallel mode similar to what we are doing for
> non-zero tuple count case. Attached patch fixes the problem.
This seems like a really ugly fix. It might be possible to come up
with a fix along these lines, but I don't have much confidence in the
specific new test you've injected into executePlan(). Broadly, any
change of this type implicitly changes the contract between
executePlan() and the planner infrastructure - the planner can now
legally generate parallel plans in some cases where that would
previously have been unacceptable. But I'm not in a hurry to rethink
where we've drawn the line there for 9.6. Let's punt this issue for
now and come back to it in a future release.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company