On Fri, Aug 31, 2018 at 3:50 PM, Andres Freund <andres@anarazel.de> wrote:
On 2018-08-31 19:31:47 -0300, Alvaro Herrera wrote: > On 2018-Aug-31, David Pacheco wrote: > > > From reading the 9.6.3 source, it looks like the autovacuum process > > itself is single-threaded, and it reads pages essentially linearly > > from the relation (possibly skipping some). When the autovacuum > > process needs to modify a page, it doesn't write it directly, but > > rather marks the buffer dirty. The page will be written later, > > Unless there's some bug, there is a BufferAccessStrategy that only lets > a few dozen buffers go unwritten before the autovac worker process > itself is forced to write some.
I've not re-checked, but I'm not sure that's true if the buffer is already in s_b, which it'll be for many workloads.
Does that mean this analysis from above is accurate?
It looks to me like the autovacuum process is effectively generating work (in the form of async writes) that's being distributed implicitly to the various backend processes, creating latency for any other query that happens to require a buffer (including read-only queries).