> That would explain all the writes, but it doesn't seem to explain why
> your two servers aren't behaving similarly.
Well, that's why I said "ostensibly identical". There may in fact be
differences, not just in the databases but in some OS libs as well.
These servers have been in production for quite a while, and the owner
has a messy deployment process.
> Most likely that's the libc implementation of the select()-based sleeps
> for vacuum_cost_delay. I'm still suspicious that the writes are eating
> more cost_delay points than you think.
Tested that. It does look like if I increase vacuum_cost_limit to 10000
and lower vacuum_cost_page_dirty to 10, it reads 5-7 pages and writes
2-3 before each pollsys. The math seems completely wrong on that,
though -- it should be 50 and 30 pages, or similar. If I can, I'll test
a vacuum without cost_delay and make sure the pollsys() are connected to
the cost delay and not something else.
-- -- Josh Berkus PostgreSQL Experts Inc.
http://www.pgexperts.com