feat: Computed field serialization for TypedDict

[michaelhly]

Change Summary

Updated computed field serialization to use the serialization function directly (if one exists) instead of always looking up a model attribute.

Related issue number

fix #657

Checklist

• Unit tests for the changes exist
• Documentation reflects the changes where applicable
• Pydantic tests pass with this pydantic-core (except for expected changes)
• My PR is ready to review, please add a comment including the phrase "please review" to assign reviewers

Selected Reviewer: @samuelcolvin

[codecov]

Codecov Report

Merging #1018 (a4f0ca1) into main (23d1065) will increase coverage by 0.03%.
The diff coverage is 100.00%.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1018      +/-   ##
==========================================
+ Coverage   93.12%   93.15%   +0.03%     
==========================================
  Files         106      106              
  Lines       15952    15994      +42     
  Branches       35       35              
==========================================
+ Hits        14855    14899      +44     
+ Misses       1090     1088       -2     
  Partials        7        7              

Files	Coverage Δ
src/serializers/computed_fields.rs	97.04% <100.00%> (+0.97%)	⬆️

... and 2 files with indirect coverage changes

---

Continue to review full report in Codecov by Sentry.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 23d1065...a4f0ca1. Read the comment docs.

[codspeed-hq]

CodSpeed Performance Report

Merging #1018 will improve performances by 13.01%

_{Comparing michaelhly:computed-field-ser-func (a4f0ca1) with main (23d1065)}

Summary

⚡ 1 improvements
✅ 139 untouched benchmarks

Benchmarks breakdown

	Benchmark	main	michaelhly:computed-field-ser-func	Change
⚡	test_generator_rust	35.3 µs	31.2 µs	+13.01%

[michaelhly]

please review

[davidhewitt]

Looks generally like a fine approach, with a couple of small thoughts. I wonder how this looks when applied to a TypedDict? I'm unsure if it's actually valid to have methods on a TypedDict?

[michaelhly]

Looks generally like a fine approach, with a couple of small thoughts. I wonder how this looks when applied to a TypedDict? I'm unsure if it's actually valid to have methods on a TypedDict?

We compute based on the serialization function provided in the schema instead of adding a method on the TypedDict model. Here is the corresponding unit test for the described case:
147e917

[adriangb]

I'm unsure if it's actually valid to have methods on a TypedDict?

At runtime yes. But type checkers hate it (insert meme reference).

[davidhewitt]

👍 so I think then in principle we can allow these fields on TypedDict and users just have to type: ignore them?

[davidhewitt]

I don't think it's fair to call this "backwards compatibility" when this is still expected to be the main code path for models and dataclasses.

I wonder if there might be a more unified way. For a model or dataclass A with computed field b, the analogous functionality really seems to be A.b.__get__(instance). For a TypedDict it looks like that also works:

>>> class Bar(TypedDict):
...      @property
...      def y(self):
...          return 434
...
>>> Bar.y.__get__({})
434

So maybe what we really want, in all cases, is

let property_value = input_value.get_type().getattr(field.property_name_py.as_ref(py))?.call_method1("__get__", (input_value,))?;

[michaelhly]

My thinking was that a serialization function should be provided to a computed field, doesn't matter if the input type is a Model, TypedDict, Dataclass, etc.

The default behavior is to compute the computed value from the function provided in the serialization schema and then it gets set in the output_dict:

pydantic-core/src/serializers/computed_fields.rs

Lines 144 to 154 in 866eb2d

	let value = self
	.serializer
	.to_python(next_value, next_include, next_exclude, extra)?;
	if extra.exclude_none && value.is_none(py) {
	return Ok(());
	}
	let key = match extra.by_alias {
	true => self.alias_py.as_ref(py),
	false => property_name_py,
	};
	output_dict.set_item(key, value)?;

This seems more generalizable to all computed fields instead of relying on the computed field defined as an attribute on the input value.

However, I am probably missing some context on how computed fields are used by https://github.com/pydantic/pydantic.

[adriangb]

I'm not following all that closely but my 2c is that ideally we extract the function from the thing in pydantic and not in pydantic-core so that:

1. We have more flexibility. It's easier to hack things (like rebuild __mro__ based on __orig_bases__ which we do for TypedDict)
2. It ensures that we do this at schema build time and not runtime

The con of that last one is that in theory someone could want us to use the method on a subclass they pass in as a value, which doesn't apply to TypedDict but also is not what we do for BaseModel and no one has complained 😄