Hi list,
I've got a heavily-updated table, and about 30 customers on the same system
each with his own version of the table. The 3 configured autovacuum workers
take turns vacuuming the table in each customer db; autovacuum is never idle
and takes a large part of the available IO.
Fearing that vacuuming might accumulate lateness and hoping to see the system
idle every now and then, I increased autovacuum_vacuum_cost_limit to 500 and
decreased autovacuum_vacuum_cost_delay to 10. First question : is it an
intelligent thing to do or am I better off ignoring the constant vacuuming and
trusting that things will get done in time ? With the new settings, autovacuum
is still constant (even though each one takes less time), but I'm wary of
making autovacuum even less "io-nice".
Second thing : the vacuumed tables+indexes taken together are bigger than the
available OS disk cache. Does vacuuming them fill the cache, or is there some
kind of O_DIRECT in use ? I have a feeling (very un-verified) that this is not
the most usefull data I could have in my cache.
This is all on PG 8.3. I know upgrading would improve things (particularly
since a large percentage of the table remains static between vacuums), but
we're still too busy for that right now (unless you tell me I'm going to see a
night-and-day difference regarding this particular issue).
Thanks.
--
Vincent de Phily
