On Tue, May 31, 2016 at 3:19 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
I wrote: > At the risk of opening another can of worms, what about renaming > max_worker_processes as well? It would be a good thing if that > had "cluster" in it somewhere, or something that indicates it's a > system-wide value not a per-session value. "max_workers_per_cluster" > would answer, though I'm not in love with it particularly.
Actually, after a bit more thought, maybe "max_background_workers" would be a good name? That matches its existing documentation more closely:
Sets the maximum number of background processes that the system can support. This parameter can only be set at server start. The default is 8.
However, that would still leave us with max_background_workers as the cluster-wide limit and max_parallel_workers as the per-query-node limit. That naming isn't doing all that much to clarify the distinction.
Node execution is serial, right?
I know work_mem has a provision about possibly using multiples of the maximum in a single query...
The per-node dynamic does seem important to at least well-document so that when users see 3 workers on one explain node and 2 on another they aren't left scratching their heads. We don't reserve a set number of workers per-statement but rather they are retrieved as needed during query execution.
I think I see the same amount of added value in tacking on 'per_cluster" as you do in adding "additional" to get "max_additional_parallel_workers" - though further refinement of trying to actively NOT consider a leader a worker would support the omission of addition[al]...
If we left max_worker_processes as defining only the maximum allowed non-parallel workers and added a new "max_cluster_workers" that would cap the combination of both "max_worker_processes" and "max_parallel_workers"...
