Re: [HACKERS] parallel.c oblivion of worker-startup failures

On Wed, Jan 24, 2018 at 10:38 AM, Peter Geoghegan <pg(at)bowt(dot)ie> wrote:
> On Tue, Jan 23, 2018 at 9:02 PM, Thomas Munro
> <thomas(dot)munro(at)enterprisedb(dot)com> wrote:
>>> Yes, this is what I am trying to explain on parallel create index
>>> thread. I think there we need to either use
>>> WaitForParallelWorkersToFinish or WaitForParallelWorkersToAttach (a
>>> new API as proposed in that thread) if we don't want to use barriers.
>>> I see a minor disadvantage in using WaitForParallelWorkersToFinish
>>> which I will say on that thread.
>>
>> Ah, I see. So if you wait for them to attach you can detect
>> unexpected dead workers (via shm_mq_receive), at the cost of having
>> the leader wasting time waiting around for forked processes to say
>> hello when it could instead be sorting tuples.
>
> The leader can go ahead and sort before calling something like a new
> WaitForParallelWorkersToAttach() function (or even
> WaitForParallelWorkersToFinish()). If we did add a
> WaitForParallelWorkersToAttach() function, then the performance hit
> would probably not be noticeable in practice. The
> parallel_leader_participates case would still work without special
> care (that's the main hazard that makes using a barrier seem
> unappealing).
>

+1. I also think so. Another advantage is that even if one of the
workers is not able to start, we don't need to return an error at the
end of the query, rather we can detect that situation early and can
execute the query successfully. OTOH, if we are not convinced about
performance implications, then WaitForParallelWorkersToFinish should
serve the purpose which can be called at later point of time.

--
With Regards,
Amit Kapila.
EnterpriseDB: http://www.enterprisedb.com