We use docker+k8s to deploy multiple rasa servers. When there are only a few rasa nodes, QPS grows roughly linearly, but when the number of nodes exceeds a certain number, QPS growth becomes very small, and adding more nodes, QPS basically does not changed. We checked that the usage of docker cores and memory did not reach the upper limit. In addition to possible I/O issues (we store the sessions in ES), what are the possible reasons for this situation?
What kind of Tracker Store / Locker Store are you using? Postgres + Redis?