Python Bytes #471 The ORM pattern of 2026?
Mar 2, 2026
A lively dive into preferring raw SQL with dataclasses for AI-friendly, typed data access and performance gains. A maintenance update on pytest-check and added mypy support. A look at dataclass-wizard for fast, typed parsing across formats. A spotlight on SQLiteo, a macOS SQLite browser, and a chat about agents, trust, and how AI shifts development workflows.
AI Snips
Chapters
Transcript
Episode notes
Raw Queries Are Better For AI Coding
- Michael Kennedy argues AI coding is far better trained on raw database queries than on specific ORMs, making raw queries more robust for agentic development.
- He measured PyMongo-style libraries having tens to hundreds of times more training footprint than individual ORMs, so AI will prefer native query patterns.
Return Dataclasses From Raw Data Access
- Do implement a small data access layer that uses raw queries internally but returns typed dataclass objects to callers for IDE support and type safety.
- Michael recommends dataclasses (or optionally Pydantic externally) to avoid heavy runtime validation in every database read.
Dataclasses Trade Validation For Performance
- Dataclasses give type hints, low overhead, and no extra dependencies, making them a lightweight alternative to ORMs or Pydantic in the data access layer.
- Michael notes Pydantic adds runtime validation cost and may be unnecessary for reads since bad data is already stored.