Thread: Autovacuum Woes

Hi All,

We have a PostgreSQL 9.2 database providing backend services for an
online multiplayer game. Three (of several) tables in this DB are
transactional tables - they in order, see 10 million, 7 million and
1.6 million new inserts daily.

Several months back, we started noticing slow inserts to these tables
(and high load average on the DB server), and eventually traced the
problem to the autovacuum process. As a workaround, we started
disabling autovacuum during our peak load times. We run from cron:
ALTER TABLE table_name SET (autovacuum_enabled = FALSE) to disable and
once our concurrent player count starts going down, set
autovacuum_enabled to TRUE to enable it.

This worked fine for many months, but of late a new problem has
started. At the time of disabling and enabling, we are seeing large
performance degradation. Several hundred connections like this:

19090 ?        Ss     0:00 postgres: user dbname 10.13.36.19(42782)
PARSE waiting
19091 ?        Ss     0:00 postgres: user dbname 10.13.36.19(42783)
PARSE waiting
19092 ?        Ss     0:00 postgres: user dbname 10.13.36.19(42784)
PARSE waiting
19093 ?        Ss     0:00 postgres: user dbname 10.13.36.19(42785)
PARSE waiting
19095 ?        Ss     0:00 postgres: user dbname 10.13.36.19(42786)
PARSE waiting

So obviously, disabling/enabling autovacuum has side-effects.

It is not clear to me what is going wrong here. What will happen if I
permanently leave autovacuum off on these tables?

The documentation has some points, but I think it is a bit confusing.
Any links to a lucid explanation of the concept, algorithm,
configuration and best practices?

TIA,

Binand

On Tue, Aug 9, 2016 at 1:02 AM, Binand Sethumadhavan <binand@gmx.net> wrote:

> Several months back, we started noticing slow inserts to these tables
> (and high load average on the DB server), and eventually traced the
> problem to the autovacuum process. As a workaround, we started
> disabling autovacuum during our peak load times. We run from cron:
> ALTER TABLE table_name SET (autovacuum_enabled = FALSE) to disable and
> once our concurrent player count starts going down, set
> autovacuum_enabled to TRUE to enable it.

I have found it safer and more effective to leave autovacuum
running at all times and add VACUUM ANALYZE commands during the off
peak hours.  This will prevent autovacuum from kicking in during
peak load periods unless there is, for example, a large number of
updates to a small table -- in which case the autovacuum run will
normally be pretty quick and will prevent a slowdown.  Of course,
you may need to run autovacuum a bit, too; but there aren't enough
clues here to suggest anything specific.

--
Kevin Grittner
EDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company

Agreed with kevin.
Also i would suggest to have a look on your pgcatalog table Of your database as they may be having bloat if temp tables are getting used too much which may cause slowness of sql performance as well.

Thanks