if you are running PG like a KV store (simple lookups/simple updates/inserts) where most of the dataset fits in memory then it quickly becomes CPU bound and ch

if you are running PG like a KV store (simple lookups/simple updates/inserts) where most of the dataset fits in memory then it quickly becomes CPU bound and changes in PG configuration/access patterns can have large effects. for example i've seen running pgbouncer on the same host drop performance by about 33%, and switching from simple protocol to extended protocol single use prepared statements drop performance by 33%.