We have an application where we are collecting timestamp-based data. Since a month of data goes over 30M rows (and growing each month), we decided to partition the data on an year-and-month basis.
We basically have one master table and partition tables like master_y2013m01 and so on. Because of our application stack we use a before insert trigger to insert rows into the right partition and in the master table and then a post insert trigger to delete them from master table (we need the original insert to return the inserted row info).
This means the master table becomes quickly populated with dead tuples and triggers the autovacuum frequently (we are using default settings for autovacuum).
After upgrading our DB from PostgreSQL 8.4 to 9.1 we noticed a large increase in IO and traced it back to auto vacuum: iotop reports bursts of 4-6M/s taking 90 seconds followed by a 30 seconds pause. That's a lot of data to chew on especially since the master table is only a couple of megabytes (indexes included).
Can anyone explain why this is happening and how we can get around it?