Skip to content

biqqles/dataclassy

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

dataclassy

dataclassy is a reimplementation of data classes in Python - an alternative to the built-in dataclasses module that avoids many of its common pitfalls. dataclassy is designed to be more flexible, less verbose, and more powerful than dataclasses, while retaining a familiar interface.

In short, dataclassy is a library for moving data around your Python programs that's optimised for speed, simplicity and developer happiness.

from dataclassy import dataclass
from typing import Dict

@dataclass
class Pet:
    name: str
    species: str
    fluffy: bool
    foods: Dict[str, int] = {}

Why use dataclassy?

This section describes various motivations for using dataclassy over dataclasses.

Backwards compatibility

dataclassy implements the decorator options of the latest version of dataclasses, plus its own, such that they are compatible back to Python 3.6. It is tested against CPython 3.6 through 3.11 and PyPy 3.6 through 3.9.

Upgrade your data classes

  • new decorator options
    • slots generates __slots__ to reduce memory footprint and improve attribute lookup performance
    • kwargs appends **kwargs to __init__'s parameter list to consume unexpected arguments
    • iter allows class instances to be destructured, like named tuples
    • hide_internals automatically hides private fields from __repr__ and excludes them from comparison and iteration
  • @dataclass usage and options are inherited (subclasses do not have to reuse the decorator)
  • fields can be in any order - fields with defaults are reordered - making inheritance feasible
  • mutable default values (list, set, dict and more) are automatically copied upon initialisation
  • new functions: is_dataclass_instance and values

Additionally, dataclassy

  • implements all the decorator options and functions of dataclasses
  • is tiny (~160 LOC; less than 25% the size of dataclasses)
  • has no external dependencies, and no stdlib imports other than types, typing and reprlib
  • is fast, matching dataclasses' performance when slots=False and significantly exceeding it when slots=True
  • supports multiple inheritance and custom metaclasses
  • comes with support for mypy

Other differences

dataclassy's API is strongly influenced by dataclasses', but with a focus on minimalism and elegance.

  • there's no field or Field. Use Hashed, Internal or factory to replicate its functions
  • there's no InitVar. Simply add arguments to __post_init__
  • there's no need for ClassVar. Simply omit the field's type hint to ignore it

Also consider

  • attrs if you need complex validation and type conversions
  • pydantic if you need strict type checking

Usage

Installation

Install the latest release from PyPI with pip:

pip install dataclassy

Or install the latest development version straight from this repository:

pip install https://github.com/biqqles/dataclassy/archive/master.zip -U

Migration

dataclassy's API is broadly similar to dataclasses. If you simply use the decorator and other functions, it is possible to instantly migrate from dataclasses to dataclassy by simply changing

from dataclasses import *

to

from dataclassy import *

Otherwise, you will have to make a couple of easy refactorings (that should leave you with cleaner code!). Consult the table under Differences or skip ahead to Examples to see dataclassy in action.

Similarities

dataclassy's dataclass decorator takes all of the same arguments as dataclasses', plus its own, and should therefore be a drop-in replacement.

dataclassy also implements all dataclasses' functions: is_dataclass, fields, replace, make_dataclass, asdict and astuple (the last two are aliased from as_dict and as_tuple respectively), and they should work as you expect.

Differences

Although dataclassy's API is similar to dataclasses', compatibility with it is not a goal (this is similar to the relationship between Crystal and Ruby).

dataclassy has several important differences from dataclasses, mainly reflective of its minimalistic style and implementation. These differences are enumerated below and fully expanded on in the next section.

dataclasses dataclassy
init-only variables fields with type InitVar arguments to __post_init__
class variables fields with type ClassVar fields without type annotation
mutable defaults a: Dict = field(default_factory=dict) a: Dict = {}
dynamic defaults b: MyClass = field(default_factory=MyClass) b: MyClass = factory(MyClass)
field excluded from repr c: int = field(repr=False) Internal type wrapper or _name
"late init" field d: int = field(init=False) d: int = None
abstract data class class Foo(ABC): class Foo(metaclass=ABCMeta):

There are a couple of minor differences, too:

  • fields returns Dict[str, Type] instead of Dict[Field, Type] and has an additional parameter which filters internal fields
  • Attempting to modify a frozen instance raises AttributeError with an explanation rather than FrozenInstanceError

Finally, there are some quality of life improvements that, while not being directly implicated in migration, will allow you to make your code cleaner:

  • @dataclass does not need to be applied to every subclass - its behaviour and options are inherited
  • Unlike dataclasses, fields with defaults do not need to follow those without them. This is particularly useful when working with subclasses, which is almost impossible with dataclasses
  • dataclassy adds a DataClass type annotation to represent variables that should be generic data class instances
  • dataclassy has the is_dataclass_instance suggested as a recipe for dataclasses built-in
  • The generated comparison methods (when order=True) are compatible with supertypes and subtypes of the class. This means that heterogeneous collections of instances with the same superclass can be sorted

It is also worth noting that internally, dataclasses and dataclassy work in different ways. You can think of dataclassy as turning your class into a different type of thing (indeed, it uses a metaclass) and dataclasses as merely adding things to your class (it does not).

Examples

The basics

To define a data class, simply apply the @dataclass decorator to a class definition (see above).

Without arguments to the decorator, the resulting class will behave very similarly to its equivalent from the built-in module. However, dataclassy's decorator has some additional options over dataclasses', and it is also inherited so that subclasses of data classes are automatically data classes too.

The decorator generates various methods for the class. Which ones exactly depend on the options to the decorator. For example, @dataclass(repr=False) will prevent a __repr__ method from being generated. @dataclass is equivalent to using the decorator with default parameters (i.e. @dataclass and @dataclass() are equivalent). Options to the decorator are detailed fully in the next section.

Class variables

You can exclude a class attribute from dataclassy's mechanisms entirely by simply defining it without a type annotation. This can be used for class variables and constants.

Parameter reordering

dataclassy modifies the order of fields when converting them into parameters for the generated __init__. Specifically, fields with default values always follow those without them. This stems from Python's requirement that parameters with default arguments follow those without them. Conceptually, you can think of the process to generate the parameter list like this:

  1. dataclassy takes the fields in definition order
  2. it splits them into two lists, the first being fields without default values and the second being fields with them
  3. it appends the second list to the first

This simple design decision prevents the dreaded TypeError: non-default argument '...' follows default argument error that anyone who has tried to do serious inheritance using dataclasses will know well.

You can verify the signature of the generated initialiser for any class using signature from the inspect module. For example, using the definition linked to above, inspect.signature(Pet) will return (name: str, species: str, fluffy: bool, foods: Dict[str, int] = {})).

If we then decided to subclass Pet to add a new field, hungry:

@dataclass
class HungryPet(Pet):
    hungry: bool

You will see that inspect.signature(HungryPet) returns (name: str, species: str, fluffy: bool, hungry: bool, foods: Dict[str, int] = {}).

Inheritance

Unlike dataclasses, dataclassy's decorator only needs to be applied once, and all subclasses will become data classes with the same options as the parent class. The decorator can still be reapplied to subclasses in order to apply new parameters.

To change the type, or to add or change the default value of a field in a subclass, simply redeclare it in the subclass.

Post-initialisation processing

If an initialiser is requested (init=True), dataclassy automatically sets the attributes of the class upon initialisation. You can define code that should run after this happens - this is called post-init processing.

The method that contains this logic should be called __post_init__. Like with dataclasses, if init=False or the class has no fields, __post_init__ will not be called.

@dataclass
class CustomInit:
    a: int
    b: int
    
    def __post_init__(self):
        self.c = self.a / self.b

In this example, when the class is instantiated with CustomInit(1, 2), the field c is calculated as 0.5.

Like with any function, your __post_init__ can also take parameters which exist only in the context of __post_init__. These can be used for arguments to the class that you do not want to store as fields. A parameter cannot have the name of a class field; this is to prevent ambiguity.

Default values

Default values for fields work exactly as default arguments to functions (and in fact this is how they are implemented), with one difference: for copyable defaults, a copy is automatically created for each class instance. This means that a new copy of the list field foods in Pet above will be created each time it is instantiated, so that appending to that attribute in one instance will not affect other instances. A "copyable default" is defined as any object implementing a copy method, which includes all the built-in mutable collections (including defaultdict).

If you want to create new instances of objects which do not have a copy method, use the factory function. This function takes any zero-argument callable. When the class is instantiated, this callable is executed to produce a default value for the field:

class MyClass:
    pass

@dataclass
class CustomDefault:
    m: MyClass = factory(MyClass)

CustomDefault()  # CustomDefault(m=<__main__.MyClass object at 0x7f8b156feb50>)
CustomDefault()  # CustomDefault(m=<__main__.MyClass object at 0x7f8b156fc7d0>)

API

Decorator

@dataclass(init=True, repr=True, eq=True, order=False, unsafe_hash=False, frozen=False, hide_internals=True, iter=False, kwargs=False, slots=False, meta=DataClassMeta)

The decorator used to signify that a class definition should become a data class. The decorator returns a new data class with generated methods as detailed below. If the class already defines a particular method, it will not be replaced with a generated one.

Without arguments, its behaviour is, superficially, almost identical to its equivalent in the built-in module. However, dataclassy's decorator only needs to be applied once, and all subclasses will become data classes with the same parameters. The decorator can still be reapplied to subclasses in order to change parameters.

A data class' fields are defined using Python's type annotations syntax. To change the type or default value of a field in a subclass, simply redeclare it.

This decorator takes advantage of two equally important features added in Python 3.6: variable annotations and dictionaries being ordered. (The latter is technically an implementation detail of Python 3.6, only becoming standardised in Python 3.7, but is the case for all current implementations of Python 3.6, i.e. CPython and PyPy.)

Decorator options

The term "field", as used in this section, refers to a class-level variable with a type annotation. For more information, see the documentation for fields() below.

init

If true (the default), generate an __init__ method that has as parameters all fields up its inheritance chain. These are ordered in definition order, with all fields with default values placed towards the end, following all fields without them. The method initialises the class by applying these parameters to the class as attributes. If defined, it will also call __post_init__ with any remaining arguments.

This ordering is an important distinction from dataclasses, where all fields are simply ordered in definition order, and is what allows dataclassy's data classes to be far more flexible in terms of inheritance.

A shallow copy will be created for mutable arguments (defined as those defining a copy method). This means that default field values that are mutable (e.g. a list) will not be mutated between instances.

repr

If true (the default), generate a __repr__ method that displays all fields (or if hide_internals is true, all fields excluding internal ones) of the data class instance and their values.

eq

If true (the default), generate an __eq__ method that compares this data class to another of the same type as if they were tuples created by as_tuple, excluding internal fields if hide_internals is true.

order

If true, a __lt__ method is generated, making the class orderable. If eq is also true, all other comparison methods are also generated. These methods compare this data class to another of the same type (or a subclass) as if they were tuples created by as_tuple, excluding internal fields if hide_internals is true. The normal rules of lexicographical comparison apply.

unsafe_hash

If true, force the generation of a __hash__ method that attempts to hash the class as if it were a tuple of its hashable fields. If unsafe_hash is false, __hash__ will only be generated if eq and frozen are both true.

match_args

If true (the default), generate a __match_args__ attribute that enables structural pattern matching on Python 3.10+.

kw_only

If true, all parameters to the generated __init__ are marked as keyword-only. This includes arguments passed through to __post_init__.

frozen

If true, instances are nominally immutable: fields cannot be overwritten or deleted after initialisation in __init__. Attempting to do so will raise an AttributeError. Warning: incurs a significant initialisation performance penalty.

hide_internals

If true (the default), internal fields are not included in the generated __repr__, comparison functions (__eq__, __lt__, etc.), or __iter__.

iter

If true, generate an __iter__ method that returns the values of the class's fields, in order of definition, noting that internal fields are excluded when hide_internals is true. This can be used to destructure a data class instance, as with a Scala case class or a Python namedtuple.

kwargs

If true, add **kwargs to the end of the parameter list for __init__. This simplifies data class instantiation from dictionaries that may have keys in addition to the fields of the data class (i.e. SomeDataClass(**some_dict)).

slots

If true, generate a __slots__ attribute for the class. This reduces the memory footprint of instances and attribute lookup overhead. However, __slots__ come with a few restrictions (for example, multiple inheritance becomes tricky) that you should be aware of.

meta

Set this parameter to use a metaclass other than dataclassy's own. This metaclass must subclass dataclassy.dataclass.DataClassMeta.

DataClassMeta is best considered less stable than the parts of the library available in the root namespace. Only use a custom metaclass if absolutely necessary.

Functions

factory(producer)

Takes a zero-argument callable and creates a factory that executes this callable to generate a default value for the field at class initialisation time.

is_dataclass(obj)

Returns True if obj is a data class as implemented in this module.

is_dataclass_instance(obj)

Returns True if obj is an instance of a data class as implemented in this module.

fields(dataclass, internals=False)

Return a dict of dataclass's fields and their types. internals selects whether to include internal fields. dataclass can be either a data class or an instance of a data class.

A field is defined as a class-level variable with a type annotation. Variables defined in the class without type annotations are completely excluded from dataclassy's consideration. Class variables and constants can therefore be indicated by the absence of type annotations.

values(dataclass, internals=False)

Return a dict of dataclass's fields and their values. internals selects whether to include internal fields. dataclass must be an instance of a data class.

as_dict(dataclass dict_factory=dict)

Recursively create a dict of a data class instance's fields and their values.

This function is recursively called on data classes, named tuples and iterables.

as_tuple(dataclass)

Recursively create a tuple of the values of a data class instance's fields, in definition order.

This function is recursively called on data classes, named tuples and iterables.

make_dataclass(name, fields, defaults, bases=(), **options)

Dynamically create a data class with name name, fields fields, default field values defaults and inheriting from bases.

replace(dataclass, **changes)

Return a new copy of dataclass with field values replaced as specified in changes.

Type hints

Internal

The Internal type wrapper marks a field as being "internal" to the data class. Fields which begin with the "internal use" idiomatic indicator _ or the private field interpreter indicator __ are automatically treated as internal fields. The Internal type wrapper therefore serves as an alternative method of indicating that a field is internal for situations where you are unable to name your fields in this way.

Hashed

Use Hashed to wrap the type annotations of fields that you want to be included in a data class' __hash__. The value hashed by __hash__ consists of a tuple of the instance's type followed by any fields marked as Hashed.

DataClass

Use this type hint to indicate that a variable, parameter or field should be a generic data class instance. For example, dataclassy uses these in the signatures of as_dict, as_tuple and values to show that these functions should be called on data class instances.

Mypy support

In order to use dataclassy in projects with mypy, you will need to use the mypy plugin. You can create a mypy.ini or .mypy.ini for such projects with the following content:

[mypy]
plugins = dataclassy.mypy