atelier/languages/python.md

Python — Language Application

How Atelier's domain principles apply in Python specifically. Derives from domains/ docs.

Type System (C1 Correctness, Data P7 Type Fidelity)

• Type hints on every function: def get_user(id: UUID) -> User | None:.
• mypy --strict or pyright in CI: type check is not optional.
• No Any without justification: Any disables the type checker. Use object + narrowing.
• Pydantic for runtime validation: schemas validate and type at the boundary.

from pydantic import BaseModel
from uuid import UUID

class UserCreate(BaseModel):
    email: str
    name: str
    # additionalProperties: false by default (extra='forbid')

Error Handling (Errors P1 Errors are Data)

• Exceptions for exceptional cases, not control flow. raise not return None for errors.
• Custom exception hierarchy:

class AppError(Exception): pass
class ValidationError(AppError): pass
class NotFoundError(AppError): pass

• Never bare except:: except Exception as e: (catch specific, not everything).
• Never except: pass: log and re-raise or handle, never swallow (Errors P2).

Async (Concurrency P7, P8)

• asyncio for I/O-bound: async def, await. Not threads for I/O.
• anyio for portability if you may switch runtimes (trio compatibility).
• Timeout on every await: asyncio.wait_for(coro, timeout=5), not bare await.
• Cancellation propagated: asyncio.CancelledError is not caught; it propagates.

Immutability (Concurrency P1)

• frozen=True dataclasses for value objects:

from dataclasses import dataclass
@dataclass(frozen=True)
class UserId:
    value: str

• Tuples over lists for fixed-length, immutable sequences.
• No in-place mutation of shared state: return new objects.

Nullability (C1)

• Optional[T] is T | None: explicit, must be checked.
• None is not "not found": raise NotFoundError or return Result, not None.
• assert is for invariants, not for runtime checks (stripped with -O).

Testing (Testing)

• pytest with fixtures (factories, not shared state).
• pytest --randomly to catch order-dependent tests (P2 Independence).
• freezegun for time: no datetime.now() in tests; inject the clock.
• factory_boy or pytest-factoryboy for realistic factories.

Observability (Observability P1)

• structlog or python-json-logger: JSON logs, not print.
• logging with structured formatter: every log has request_id, user_id, event.
• No secrets in logs: mask_secret() helper, or structlog processors.

Tooling (DevOps P2)

• ruff for lint + format: replaces flake8 + black + isort.
• mypy --strict in CI: type check.
• pip-tools or poetry for lockfile: pinned dependencies.
• pip install --no-deps -r requirements.txt: reproducible install.