[fhdsgbbcaA]

Seems like the issue is less using dicts than not treating external APIs as input that needs to be sanitized.

[physicsguy]

The code in the examples doesn't even check the API response code, let alone the structure of the response.

[pmarreck]

Agreed. If you sanitize/allowlist API data you should not have issues with dicts.

[imron]

You'll have issues if you ever rename things in the dict.

Linting tools will pick up on every instance where you forgot to rename the fields of a class, but won't do the same for dicts.

[FreakLegion]

TypedDicts solve the linting problem, but refactoring tools haven't caught up (unlike e.g. ForwardRef type annotations, which are strings but can be transformed alongside type literals).

[tomjakubowski]

Is there any advantage to using a TypedDict for a record over a dataclass?

[FreakLegion]

TypedDicts "aren't real" in the sense that they're a compile-time feature, so you're getting typing without any deserialization cost beyond the original JSON. Dataclasses and Pydantic models are slow to construct, so that's not nothing.

This of course means TypeDicts don't give you run-time validation. For that, and for full-blown custom types in general, I tend to favor msgspec Structs: https://jcristharif.com/msgspec/benchmarks.html#json-seriali....

[orf]

> Dataclasses and Pydantic models are slow to construct

Citation needed? Pydantic is really quite fast, and you can pass raw JSON responses into it.

It may be slower (depending on the validators or structure), but I’d expect it to be comparably fast to the stdlib JSON module.