(2019/03/11 13:06), Tom Lane wrote:
> Etsuro Fujita<fujita.etsuro@lab.ntt.co.jp> writes:
>>>> The parallel safety of the final scan/join target is determined from the
>>>> grouping target, not that target, which [ is wrong ]
>
>> This would only affect plan quality a little bit, so I don't think we
>> need a regression test for this, either, but the fix might destabilize
>> existing plan choices, so we should apply it to HEAD only?
>
> Is that the only possible outcome? Per Robert's summary quoted above,
> it seems like it might be possible for the code to decide that the final
> scan/join target to be parallel safe when it is not, leading to outright
> wrong answers or query failures.
Maybe I'm missing something, but I think that if the final scan/join
target is not parallel-safe, then the grouping target would not be
parallel-safe either, by the construction of the two targets, so I don't
think that we have such a correctness issue.
Best regards,
Etsuro Fujita