the SQLAlchemy connection pool (which is what this is using) calls connection.rollback() when the connection is checked back in. as far as "leaking" transactional state it's not different from "autocommit", which is what Postgresql connections normally do (and what psycopg2 will do if you set `connection.autocommit=True`).
How is a connection "checked back in"?
Aside from the severe risk of resource leaks, that has implications for which pooling modes you can use. Particularly, you can't use statement-level pooling, and unless you're scrupulous about managing those implicitly created transactions, transaction-level pooling is pretty much the same thing as connection-level pooling.
Adjust your expectations accordingly, or turn on auto-commit.
(EDIT: clarification.)