A namedtuple-style library for defining immutable sum types in Python. (Get it on PyPI)

You may know sum types under a different name – they're also referred to as tagged unions, enums in Rust/Swift, and variants in C++. If you haven't heard about them yet, here's a nice introduction.

The current stable version is 0.10.0. The library supports Python 3.5+. The core code has lived in various utils folders for about a year, before I got tired of copying it around and decided to release it as an independent package. (see also: Should I use it?)

Suggestions, feedback and contributions are very welcome!

>>> from sumtype import sumtype
>>> from typing import Tuple
>>>
>>> class Thing(sumtype):
...     def Foo(x: int, y: int): ...
...     def Bar(y: str, hmm: Tuple[str, str]): ...
...     def Zap(): ...
...
>>>

This means that a Thing value can be one of three variants:

• a Foo with two int fields, x and y
• a Bar with a str field y and a Tuple[str, str] field hmm
• a Zap with no fields

If type annotations are provided, the constructors will typecheck the arguments (see Typechecking) You can also add your own docstring and methods in the class definition. If you prefer namedtuple-style definitions, sumtype supports those too - see Thing2 in sumtype.sumtype.demo() for an example.

>>> foo = Thing.Foo(x=3, y=5)  # named arguments
>>> bar = Thing.Bar('hello', ('wo', 'rld'))  # positional arguments
>>> zap = Thing.Zap()

Note that they're still just different values of the same type, not subclasses:

>>> type(foo) is Thing  and  type(bar) is Thing  and  type(zap) is Thing
True

Every specified field gets an accessor:

>>> foo.x, foo.y;
(3, 5)
>>> bar.y,  bar.hmm
('hello', ('wo', 'rld'))

...with checks if the access is valid and descriptive error messages:

>>> Thing.Zap().hmm  # only `Bar`s have a `hmm` field
Traceback (most recent call last):
  ...
AttributeError: Incorrect 'Thing' variant: Field 'hmm' not declared in variant 'Zap'...
>>>
>>> Thing.Foo(x=1, y=2).blah_blah_blah  # no variant has a `blah_blah_blah` field 
Traceback (most recent call last):
  ...
AttributeError: Unknown attribute: Field 'blah_blah_blah' not declared in any variant of 'Thing'...

The values also have a nice __repr__():

>>> foo; bar; zap
Thing.Foo(x=3, y=5)
Thing.Bar(y='hello', hmm=('wo', 'rld'))
Thing.Zap()

The library is designed with efficiency in mind¹ – it uses __slots__ for attribute storage and generates specialized versions of all the methods for each class. To see the generated code, do class Thing(sumtype, verbose=True):.

¹ At least I like to think so ;) I try to do my best with profiling things though!

sumtype uses typeguard to typecheck the fields:

>>> # Foo(x: int, y: int) -> Thing
>>> Thing.Foo(x=1, y=2)
Thing.Foo(x=1, y=2)
>>> Thing.Foo(x='should be an int', y=2)
Traceback (most recent call last):
  ...
TypeError: type of argument "x" must be int; got str instead

typing annotations are supported too:

>>> # Bar(y: str, hmm: Tuple[str, str]) -> Thing
>>> Thing.Bar(y='a', hmm=('b', 'c'))
Thing.Bar(y='a', hmm=('b', 'c'))
>>> Thing.Bar(y='a', hmm=(1, 2))
Traceback (most recent call last):
  ...
TypeError: type of argument "hmm"[0] must be str; got int instead

typeguard supports all typing constructs (Tuple, List, Dict, Union, etc). (See their README for a full list) However, as of 2.2.2 it doesn't support user-defined generic classes, so for a field like z: UserDefinedList[float], typeguard will not check if the contents are actually floats. This also prevents us from defining generic sumtypes (e.g. ConsList[A], Maybe[A], Either[A, B]), but I'm working on resolving this issue.

Typechecking can be controlled with the typecheck argument: class Thing(sumtype, typecheck='always'|'debug'|'never'):. The default mode is 'always' Fields with no annotations will not be typechecked, and you can mix annotated and non-annotated fields in a definition.

>>> Thing.Foo(1,2) == Thing.Foo(1,2)
True
>>> Thing.Foo(1,2) == Thing.Bar('a', ('b', 'c'));
False
>>> {foo, foo, bar, zap} == {foo, bar, zap}
True

__eq__ and __hash__ pay attention to variant - even if we had a variant Moo(x: int, y: int), Foo(1,2) != Moo(1,2) and hash(Foo(1,2)) != hash(Moo(1,2)).

Note: Just like tuples, sumtypes __eq__/__hash__ work by __eq__ing/__hash__ing the values inside, so the values must all implement the relevant method for it to work.

>>> foo;  foo.replace(x=99)
Thing.Foo(x=3, y=5)
Thing.Foo(x=99, y=5)
>>>
>>> bar;  bar.replace(y='abc', hmm=('d', 'e'))
Thing.Bar(y='hello', hmm=('wo', 'rld'))
Thing.Bar(y='abc', hmm=('d', 'e'))

foo.replace(x=99) returns a new value, just like in namedtuple. .replace behaves just like the constructors w.r.t. typechecking.

Note: replace and all the other methods have underscored aliases (_replace). So even if you have a field called replace, you can still use my_value._replace(x=15).

>>> def do_something(val: Thing):
...     if val.is_Foo():
...         print(val.x * val.y)
...     elif val.is_Bar():
...         print('The result is', val.y, ''.join(val.hmm))
...     elif val.is_Zap():
...         print('Whoosh!')
...     else: val.impossible() # throws an error - nice if you like having all cases covered
...
>>> for val in (foo, bar, zap):
...     do_something(val)
...
15
The result is hello world
Whoosh!

>>> [ val.match(
...      Foo = lambda x, y: x*y, 
...      Zap = lambda: 999,
...      _   = lambda: -1 # default case
...   )
...  for val in (foo, bar, zap)]
[15, -1, 999]

To...

>>> foo.values();  foo.values_dict();
(3, 5)
OrderedDict([('x', 3), ('y', 5)])

>>> foo.as_tuple();  foo.as_dict()
('Foo', 3, 5)
OrderedDict([('variant', 'Foo'), ('x', 3), ('y', 5)])

...and from

>>> Thing.from_tuple(('Foo', 10, 15));  Thing.from_dict({'variant':'Foo', 'x': 10, 'y': 15})
Thing.Foo(x=10, y=15)
Thing.Foo(x=10, y=15)

Also, x == Thing.from_tuple(x.as_tuple()) and x == Thing.from_dict(x.as_dict()).

>>> import pickle
>>> vals  = [Thing.Foo(1, 2), Thing.Bar('one', ('two', 'three')), Thing.Zap()]
>>> vals2 = pickle.loads(pickle.dumps(vals))
>>> vals;  vals == vals2
[Thing.Foo(x=1, y=2), Thing.Bar(y='one', hmm=('two', 'three')), Thing.Zap()]
True

There's also tests in sumtype.tests to ensure that it all works correctly. And that's everything... for now!

• Defining generic sumtypes like Maybe[A]/Either[A, B] in a typesafe way

Yeah! I didn't just build this library because I thought it'd be nice – I'm using it heavily in an app I'm developing and a few smaller projects. Saying that it's battle-tested is a bit much, but it's getting there.

• Default values

• mypy support. Unfortunately, last time I checked, mypy didn't handle metaclass-created classes too well, but that might have changed. Alternatively, we could provide a way to generate mypy stub files. Also, right now there's no way to tell mypy that the return type of accessors depend on the variant – Union[a, b] is close, but mypy will complain that not all cases are handled even if they are.

• Statically generating a class definition to a file

• Dynamic alternatives to custom-generated methods – might be useful if startup time is more important than efficiency

• An alternative implementation backed by tuples if true immutability is desired – there's currently no way to make a __slots__-based implementation watertight in that aspect, I'm doing my best.

• Maybe opt-in mutability – currently, you can use Thing._unsafe_set_Foo_x(foo, 10) if you want that, but that's not a nice interface