Aren't you still capping the throughput by the query rate of your connection pool though? By limiting that, you are limiting the application as a whole - i.e. your benchmark is bound by the speed of your database, and has (almost) nothing to do with the performance of a specific python implementation.