Re: Sub-millisecond [autovacuum_]vacuum_cost_delay broken

Greetings,

* Thomas Munro (thomas.munro@gmail.com) wrote:
> On Fri, Mar 10, 2023 at 10:26 AM Melanie Plageman
> <melanieplageman@gmail.com> wrote:
> > I think that 4753ef37e0ed undid the work caf626b2c did to support
> > sub-millisecond delays for vacuum and autovacuum.
> >
> > After 4753ef37e0ed, vacuum_delay_point()'s local variable msec is a
> > double which, after being passed to WaitLatch() as timeout, which is a
> > long, ends up being 0, so we don't end up waiting AFAICT.
> >
> > When I set [autovacuum_]vacuum_delay_point to 0.5, SHOW will report that
> > it is 500us, but WaitLatch() is still getting 0 as timeout.
> 
> Given that some of the clunkier underlying kernel primitives have
> milliseconds in their interface, I don't think it would be possible to
> make a usec-based variant of WaitEventSetWait() that works everywhere.
> Could it possibly make sense to do something that accumulates the
> error, so if you're using 0.5 then every second vacuum_delay_point()
> waits for 1ms?

Hmm.  That generally makes sense to me.. though isn't exactly the same.
Still, I wouldn't want to go back to purely pg_usleep() as that has the
other downsides mentioned.

Perhaps if the delay is sub-millisecond, explicitly do the WaitLatch()
with zero but also do the pg_usleep()?  That's doing a fair bit of work
beyond just sleeping, but it also means we shouldn't miss out on the
postmaster going away or similar..

Thanks,

Stephen