Hi,
On 2024-06-18 16:09:09 -0500, Nathan Bossart wrote:
> On Tue, Jun 18, 2024 at 01:43:34PM -0700, Andres Freund wrote:
> > I just don't see much point in reserving 256 worker "possibilities", tbh. I
> > can't think of any practical system where it makes sense to use this much (nor
> > do I think it's going to be reasonable in the next 10 years) and it's just
> > going to waste memory and startup time for everyone.
>
> Given this, here are some options I see for moving this forward:
>
> * lower the cap to, say, 64 or 32
> * exclude autovacuum worker slots from computing number of locks, etc.
That seems good regardless
> * make the cap configurable and default it to something low (e.g., 8)
Another one:
Have a general cap of 64, but additionally limit it to something like
max(1, min(WORKER_CAP, max_connections / 4))
so that cases like tap tests don't end up allocating vastly more worker slots
than actual connection slots.
> My intent with a reserved set of 256 slots was to prevent users from
> needing to deal with two GUCs. For all practical purposes, it would be
> possible to change autovacuum_max_workers whenever you want. But if the
> extra resource requirements are too much of a tax, I'm content to change
> course.
Approximately tripling shared memory usage for tap test instances does seem
too much to me.
Greetings,
Andres Freund