A tiny one-file ORM (I checked, the full ORM itself is ultimately one file "peewee.py") like Peewee that loads objects with a super-shallow stack depth will hav

A tiny one-file ORM (I checked, the full ORM itself is ultimately one file "peewee.py") like Peewee that loads objects with a super-shallow stack depth will have very low method call overhead for the things it does do, and in Python method calls are everything. So an ORM like SQLA will appear slower on simplistic benches which compare database sequences on a near one-to-one basis. Where we try to be the most performant in the real world is that we try to make it extremely easy to cut down the number of SQL calls in a complex graph dramatically via collections, eager loading, stateful sessions, delayed flush, etc.