How do you properly clean up database state between pytest tests without killing speed?

Our API test suite uses a real PostgreSQL database (Docker, spun up in CI). Each test creates data and we need isolation. Current approach: truncate all tables in a session-scoped fixture after each test module. It works but is slow and sometimes leaves foreign key violations. Options I'm considering: 1. Wrap each test in a transaction and roll back — cleanest but some tests do multi-connection operations 2. Use `pytest-postgresql` to spin up a fresh DB per test — too slow 3. Truncate only the tables each test touches — complex to maintain What's your pattern? We're using SQLAlchemy + alembic.