Serialization with when_used=always or when_used=json?

[futurewasfree]

Hello,
I would like to understood reasoning behind design choice of when_used=json for default serialiser setup?

Pros for when_used=always from my side:

1. Consistency: model_dump() and model_dump_json() gives predictable output.
2. Default setting in pydantic itself? If you dont specify it, it's always.
3. No need of special handling/branching for two cases (corollary from 1) in client code.

It has cons too:

1. Payload size
2. Maybe reduces flexibility, depends on what you want to do next with model_dump() output..

Ideally, I guess, you want to make it configurable..

[caniko]

Ideally, I guess, you want to make it configurable..

Indeed, we should make it configurable for those that want to tinker.

Consistency: model_dump() and model_dump_json() gives predictable output.

I disagree, they are different methods, doesn't mean they should have the same effect.

Default setting in pydantic itself? If you dont specify it, it's always.

Having always as default makes sense considering most cases; doesn't mean deviating from the common norm is wrong. I agree that we should make the behavior more clear. Converting the numpy array to a data dict is an unnecessary step when we are still in a Python environment, the NumPy array stores this information already.

No need of special handling/branching for two cases (corollary from 1) in client code.

I don't understand.

Best solution from my point of view is to introduce factory methods for JSON and dict, #28. That way you can change the default behavior when desired.

Another solution is to generate code where we have all the combinations available.

[futurewasfree]

I'll look into it.. so essentially what we want is two more args to factory function with meaningful (NumpyDataDict) defaults?

[caniko]

Meaningful might be the wrong descriptor; defaults that keep the current behavior. We have to avoid breaking changes.