Thanks for your answers! I'm using 8.1 and 8.2 on windows2003 servers, and it's true that i could probably configure them much better. We've recently moved to brand new dedicated database servers with pg8.3 on debian in 2 projects and it has been much easier to configure these correctly. There I don't encounter the probems that i described.
The thing is that the whole concept of autovacuum is not feeling right. Per design, the vacuum is likely to kick off when i am doing something big. And when i am doing something big, a vacuum is the last thing i'd wish for. I'd wish for a vacuum when the database is doing nothing at all, but the autovacuum will NEVER kick off in such a moment.
That's why i feel better scheduling the vacuum at times at which i know things will be generally quiet.
To be honest, i am a bit surprised that all 3 reactions recommend using autovacuum, even if it means i have to buy a new server for this purpouse. I was thinking that autovacuum was just a mechanism to ensure that postgres works well out of the box, but that it would be recommended to schedule your own vacuum tailored to your specific needs. I agree though, that it is a tough tailoring job and that the autovacuum must be doing a better job than i am. It just fires at the wrong time.
Just a thought (to think positively..): wouldn't it be possible to let the autovacuum wait until the load goes down, or until the end of the transaction that triggered the autovacuum?
