hill.dataclassio 0.3.0

A lightweight package for converting dataclass hierarchies to and from dictionaries.

DataclassIO

DataclassIO is a package for fast roundtrip conversion between python literals and dataclass hierarchies.

This project is in its early stages. If you need maturity and lots of bells of whistles, consider another library :)

Supported Types and Constructs

Currently, the following types are supported:

• Optional expressions: Optional[T], T | None, or Union[T, None]
• Embedded dataclasses: DataclassType (exports as dict)
  • Unions of multiple dataclasses are supported via a discriminator.
• Lists: list[T]
• Dicts: dict[TK, TV]
• Enums: Enum (exports as str)
• Datetimes: datetime (imports/exports to ISO-8601 string)
• fundamental types: e.g., int, float, str, bool

Where the T, TK, and TV type variables may be any other type listed in the table. For instance, dict[str, list[DataclassType | None]] is supported.

The library also supports common dataclass features:

• Default values and default factories: field(default=..., default_factory=...)
• Init-only and init=False fields
  • Note that init=False fields are still exported.
• kw_only fields and dataclasses.
• Class variables.

The library also supports a handful of hooks that can be used to customize the serialization or deserialization process, or run other side effects during to_dict and from_dict calls.

Example

from dataclasses import dataclass, field

from dataclassio import IOMixin, EFS

@dataclass
class Address:
    city: str
    zip_code: tp.Optional[str] = None


@dataclass
class User(IOMixin):
    id: int
    name: str
    is_admin: bool = False
    address: tp.Optional[Address] = None  # <-- Address does not derive from IOMixin!
    named_addresses: dict[str, Address] = field(default_factory=dict)

## Dictionary -> Dataclass

data = {
    "id": 1,
    "name": "Alice",
    "named_addresses": {"home": {"city": "Anytown"}, "work": {"city": "Coolsville"}},
}
actual = User.from_dict(data)
assert isinstance(actual.named_addresses, dict)
assert isinstance(actual.named_addresses["home"], Address)
assert actual.named_addresses["work"].city == "Coolsville"

## Dataclass -> Dictionary (note that the default-valued-fields are excluded)
dikt = actual.to_dict(skip_defaults=True)
assert dikt == data

## Extra Fields are captured:
address_data = {
    "city": "Anytown",
    "bonus": "extra",
}
address_obj = Address.from_dict(address_data, extra_field_strategy=EFS.CAPTURE)
assert address_obj.extra_fields == {"bonus": "extra"}
assert address_obj.to_dict() == address_data  # extra fields survive round trip.

Configuration

A handful of configuration options are supported to customize serialization and deserialization. These can be provided either at time of a to_dict or from_dict call ("call-level config") or can be applied to individual dataclass fields statically ("field-level config"). Not all options can be provided in both places. The following options are supported:

• disable_hooks: A flag to skip to_dict and from_dict hooks, even if they are defined on a type.
• discriminator: The name of a field to use when discrimating between unions of dataclass types.
• extra_field_strategy: The preferred method for handling unexpected fields during deserialization (from_dict). By default they are ignored. Alternatively, an exception can be raised (STRICT-mode) or they can be captured into a private attribute of the object (CAPTURE-mode). In CAPTURE-mode, the unexpected fields are yielded in the to_dict call, supporting full round-tripping.
• include_src_in_docstring: A flag to include to_dict and from_dict method source code in their corresponding docstrings. This is useful since their source code is dynamically generated.
• skip: Field-level flag to omit a field from to_dict.
• skip_extras: Call-level flag to omit any cpatured extra fields from to_dict. If you know that your instances do not have extras, or don't care to preserve this, this can be safely toggled to True.
• skip_if_default/skip_defaults: Flags to omit fields during the to_dict call if they have the default value. skip_if_default is field-level only, and has the highest precedence. skip_defaults can be applied at the call- and field-level. At the field-level, skip_defaults has a slightly different meaning from skip_if_default:
  • x: T = field(..., FieldOptions(skip_if_default=True)): Suppress serialization of the field x if its value equals the default.
  • x: T = field(..., FieldOptions(skip_defaults=True)): Suppress serialization of T's fields if they have a default value.

Propagation and Precedence

When a configuration option is specified at both the call- and field-level, the field-level option is preferred.

In addition to precedence differences, options propagate differently depending on where they are defined. Call-level config options always propagate through fields which are themselves dataclass-typed. Field-level config options do not.

Field-level options can be divided into "local" and "deep once" propagation categories. "Local" options (e.g., discriminator) affect the code-generation for their contained dataclass but not the nested dataclass. Meanwhile, "deep once" options (e.g., extra_field_strategy) are the inverse - they do not affect the code-generation for their contained dataclass but do for the affected field. Most of the time, this distinction is the natural interpretation and does not need to be top of mind.

Benchmarks

Dict -> Dataclass

We compare the deserialization (i.e., from_dict) methods for dataclassio and similar python libraries using the COCO2017 validation dataset (~20MB; 5000 images; 36781 annotations), which is a large but shallow JSON hierarchy. For benchmarking, the JSON file is first loaded and converted to python builtins. This dataset has no extra fields, so the equivalent of extra_field_strategy=EFS.IGNORE is utilized for all libraries.

We obtain the following times using pybench run benchmarks/compare_with_other_libraries.py --profile thorough:

benchmark	time (avg)	p99	vs base
dataclassio ★	48.39 ms	50.26 ms	baseline
pydantic v2	85.82 ms	93.31 ms	1.77× slower
msgspec_into_dataclass	39.22 ms	42.99 ms	1.23× faster
msgspec_into_native	19.97 ms	20.80 ms	2.42× faster
mashumaro	62.42 ms	63.09 ms	1.29× slower
dataclass-wizard	63.40 ms	65.95 ms	1.31× slower

Only msgspec is significantly faster dataclassio, but it does not support capturing extra fields. As a key distinction, dataclassio does not try to validate or coerce types, which may explain some or all of the performance advantages relative to other libraries, which may be coercing by default. Other libraries are also more configurable, may support more types, and so forth. However, if all you need is fast dict -> dataclass conversion that supports capturing unexpected fields, then dataclassio may be a good fit.

Dataclass -> Dict

Using the same large COCO2017 dataset, we also evaluate conversion back to python literals (e.g., to_dict) using the default settings for all libraries.

We obtain the following times using pybench run benchmarks/compare_to_dict.py --profile smoke:

benchmark	time (avg)	p99	vs base
dataclassio ★	28.25 ms	28.30 ms	baseline
native	376.97 ms	377.70 ms	13.34× slower
pydantic	130.21 ms	130.70 ms	4.61× slower
msgspec_from_dataclass	92.68 ms	93.50 ms	3.28× slower
msgspec_from_struct	86.26 ms	87.91 ms	3.05× slower
mashumaro	29.74 ms	30.14 ms	1.05× slower
dataclass-wizard	33.81 ms	35.12 ms	1.20× slower

In this test, dataclassio was the fastest library available, though mashumaro is probably equivalent.