Robert Haas wrote:
> I think this approach is actually better anyway. There's no guarantee
> that VACUUM can be responsive enough to get the job done in time, work
> items or no work items,
Yeah, autovacuum work items don't have a guaranteed response time.
They're okay for things that "ought to be done eventually", but if the
condition causes a high-speed load to fail with errors, then halting the
load until the cleanup is done seems like the way to go. Having to
randomly inject pauses in your workload so that autovacuum has time to
cope isn't great from the user's POV (a condition that is made worse if
you don't have any mechanism to detect that you need a pause, or how
long to pause for.)
Once that part is working well, you could as a convenience (to avoid or
reduce the stalls some of the time) add autovacuum work-item support.
--
Álvaro Herrera https://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services