On 05/31/2016 11:27 AM, Peter Geoghegan wrote:
> On Tue, May 31, 2016 at 11:22 AM, Josh berkus <josh@agliodbs.com> wrote:
>>> I think we can hope that developers are going to be less confused about
>>> that than users.
>>
>> Makes sense.
>
> Maybe EXPLAIN doesn't have to use the term parallel worker at all. It
> can instead use a slightly broader terminology, possibly including the
> term "core".
>
>> One more consistency question: what's the effect of running out of
>> max_parallel_workers?
>>
>> That is, say max_parallel_workers is set to 10, and 8 are already
>> allocated. If I ask for max_parallel_X = 4, how many cores to I use?
>
> Well, it depends on the planner, of course. But when constrained only
> by the availability of worker processes, then your example could use 3
> cores -- the 2 remaining parallel workers, plus the leader itself.
>
>> Presumably the leader isn't counted towards max_parallel_workers?
^^^^ (oops, I meant max_worker_processes above)
So, there's six things we'd like to make consistent to limit user confusion:
1. max_parallel_X and number of cores used
2. max_parallel_X and EXPLAIN output
3. max_parallel_X and gatekeeping via max_worker_processes
4. max_parallel_X and parallelism of operations (i.e. 2 == 2 parallel
scanners)
5. settings in other similar databases (does someone have specific
citations for this)?
6. consistency with other GUCs (0 or -1 to disable settings)
We can't actually make all five things consistent, as some of them are
contradictory; for example, (1) and (3) cannot be reconciled. So we
need to evaluate which things are more likely to cause confusion.
My general evaluation would be to make the GUC be the number of
additional workers used, not the total number (including the leader).
From my perspective, that makes (2), (3) and (6) consistent, and (4)
cannot be made consistent because different types of parallel nodes
behave different ways (i.e, some are parallel with only 1 additional
worker and some are not).
However, I'm resigned to the fact that user confusion is inevitable
whichever way we choose, and could be persuaded the other way.
--
--
Josh Berkus
Red Hat OSAS
(any opinions are my own)
