I’ve posted one of these before but this one is for our master database.
The initial drop is us starting to hit pgbouncer on the server, rather than via lots of pgbouncers on the client, and switching to transaction pooling.
It got a bit messy as it turns out our server_lifetime and server_idle_timeout were tuned way down; to seconds rather than minutes, so a lot of connection churn was occuring. The pleasant flat bit at the end is our current state.
Posted by lozzd on 2011-09-23 14:35:49
Tagged: , pgbouncer , database , postgres