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

On Mon, Jan 22, 2018 at 7:16 PM, Robert Haas <robertmhaas(at)gmail(dot)com> wrote:
> On Fri, Jan 19, 2018 at 10:59 PM, Amit Kapila <amit(dot)kapila16(at)gmail(dot)com> wrote:
>> The patch doesn't apply cleanly on the head, but after rebasing it, I
>> have reviewed and tested it and it seems to be working fine. Apart
>> from this specific issue, I think we should consider making
>> nworkers_launched reliable (at the very least for cases where it
>> matters). You seem to be of opinion that can be a source of subtle
>> bugs which I don't deny but now I think we are starting to see some
>> use cases of such a mechanism as indicated by Peter G. in parallel
>> create index thread. Even, if we find some way for parallel create
>> index to not rely on that assumption, I won't be surprised if some
>> future parallelism work would have such a requirement.
>
> Isn't making nworkers_launched reliable exactly what this patch does?
> It converts the rare cases in which nworkers_launched would have been
> unreliable into errors, precisely so that code like parallel CREATE
> INDEX doesn't need to worry about the case where it's inaccurate.
>

It does turn such cases into error but only at the end when someone
calls WaitForParallelWorkersToFinish. If one tries to rely on it at
any other time, it won't work as I think is the case for Parallel
Create Index where Peter G. is trying to use it differently. I am not
100% sure whether Parallel Create index has this symptom as I have not
tried it with that patch, but I and Peter are trying to figure that
out.

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