P28 · Objects & Data Modelling

A dataclass is not a shorter way to write __init__. Used deliberately, it is where you decide which objects are allowed to exist at all — and that decision is worth more than any amount of downstream validation, because a check you remember to perform is a check someone will eventually forget. This pillar is about designing types whose invalid states cannot be constructed, and contracts that a checker rather than a convention holds callers to.

The first track models the data. Dataclasses with the options that matter — frozen for immutability where mutation buys nothing, ordering and comparison where they are meaningful, defaults that do not silently share state — and validation placed at construction, so an instance that exists is an instance you can trust. The second track handles the interface between components, and it turns on a distinction Python supports in both directions. Structural typing says anything with the right shape qualifies, which is what a Protocol expresses and what lets you accept objects from code you do not control, tests included. Nominal typing says a class must declare its intent by inheriting, which is what an abstract base class expresses and what lets you refuse accidental conformance. Neither is the default answer; the track is about knowing which question you are in.