python
Allow subclassing without supertyping
This issue comes from python/mypy#1237. I'll try to make a summary of the discussion here
Some one reported an issue with an artificial example (https://github.com/python/mypy/issues/1237#issue-136162181):
class Foo:
def factory(self) -> str:
return 'Hello'
class Bar(Foo):
def factory(self) -> int:
return 10
and then with the following stub files (https://github.com/python/mypy/issues/1237#issuecomment-188710925):
class QPixmap(QPaintDevice):
def swap(self, other: 'QPixmap') -> None: ...
class QBitmap(QPixmap):
def swap(self, other: 'QBitmap') -> None: ...
Which mypy currently reports as erroneous: Argument 1 of "swap" incompatible with supertype "QPixmap"
These were initially was argued to be a correct error because they violate Liskov Substitution Principle (the same case by a change of return type, the second is a covariant argument change). The problem in this scenario is that the actual classes are not meant to be substituted (the first example is actually a wrapper for a non-virtual C++ function so they're not substitutable, and the second aren't supposed to be mixed together, just reuse part of the code). (see https://github.com/python/mypy/issues/1237#issuecomment-189490903)
There was also a suggestion that allow covariant args explicitly (in the same way that Eiffel allows it and adds a runtime check) with a decorator
class QBitmap(QPixmap):
@covariant_args
def swap(self, other: 'QBitmap') -> None: ...
My proposal instead was add what some dialects of Eiffel call "non-conforming inheritance" ("conforms" is the Eiffel word for what PEP-483 calls "is-consistent-with"). It is essentially a mechanism to use subclassing just as a way to get the implementation from the superclass without creating a subtyping relation between them. See https://github.com/python/mypy/issues/1237#issuecomment-231199419 for a detailed explanation.
The proposal is to haven Implementation in typing so you can write:
class QBitmap(Implementation[QPixmap]):
def swap(self, other: 'QBitmap') -> None: ...
which just defines a QBitmap class with all the mothods copied from QPixmap, but without making one a subtype of the other. In runtime we can just make Implementation[QPixmap] == QPixmap
As a follow-up to the discussion, @gvanrossum suggested using a class decorator. The problem with this approach is that class-level is not the right granularity to decide this, but each individual inheritance relation is. Ina multiple-inheritance context, you could want to inherit the interface of a SuperClassA, and the implementation of SuperClassB. I think that might be quite common with mixins (actually I think most mixins could be used with this form of inheritance)
Looking deeper into what Eiffel does (but this is a bit of a side-comment, not part of the main proposal), there it was also possible to define a superclass in a way that only implementation inheritance from it was allowed, and that could work well with mixins (and implemented as a class decorator). However you still need to be able to say "I just want the implementation of this superclass" when inheriting, because the writer of the superclass may not have foreseen that you wanted that.
Also note that this idea makes self types more viable. Things like TotalOrderingMixin where you have
def __lt__(self: T, other:T): -> bool ...
can be defined more accurately (currently this is in the stdlib), and with no errors (note that the < operator/ __lt__ method is actually covariant in its argument in Python 3). TotalOrderingMixin is something that would create an error in every case with the current mypy implementation if self types are added as is.
I think this is a very nice feature. In general, it is good that we separate classes (code reuse) and types (debugging/documentation). If it is added, then I would also mention this feature in PEP483 (theory) as an example of differentiating classes and types.
Between a decorator and a special class I chose the second one, since as @dmoisset mentioned, it is more flexible and suitable fox mixin use cases. Also, it is very in the spirit of gradual typing, one can mark particular bases as Implementation while still typecheck with respect to other bases.
I'd like to explore this further. I expect that specifying semantics that work for every use case will be tricky. E.g. when using implementation inheritance, how much control does the subclass have and how much control does the superclass have? Do you have to repeat the signature for every method? Only describe exceptions? Which use cases should be care about most?
I hope that one of you can come up with a concrete, detailed proposal.
I think it is better to have a minimalistic solution, rather than try to make a complex proposal that will cover all use cases. Later, more features for most popular use cases could be added.
Here is a detailed proposal with examples:
- At runtime
Implementation[MyClass] is MyClass. So that the separation of control between subclass/superclass is as usual. Also, all normal mechanisms work at runtime:
class Base:
def __init__(self, number: int) -> None:
self.number = number
class AnotherBase:
...
class Derived(Implementation[Base], AnotherBase):
def __init__(self) -> None:
super().__init__(42) # This works since Implementation[Base] is just Base
-
Use cases for
Implementationare situations where a user wants to reuse some code in another class without creating excessive "type contracts". Of course one can useAnyor#type: ignore, butImplementationwill allow more precise type checking. -
The derived class is not considered a subtype of implementation base classes. An explicit cast might be used under user's responsibility. Examples:
from typing import cast
def func(x: Base) -> None:
...
def another_func(x: AnotherBase) -> None:
...
func(Base(42)) # this is OK
another_func(Derived()) # this is also OK
func(Derived()) # Marked as error by a typechecker
func(cast(Base, Derived())) # Passes typecheck
- Methods of an implementation base class could be redefined in an incompatible way:
class Base:
def method(self, x: int) -> None:
...
class AnotherBase:
def another_method(self, y: int) -> None:
...
class Derived(Implementation[Base], AnotherBase):
def another_method(self, y: str) -> None: ... # This is prohibited by a typechecker
def method(self, x: str) -> None: ... # Although, this is OK
- As @JukkaL proposed, a typechecker checks that the redefined methods do not "spoil" the already checked code by re-type-checking the bodies of all base class methods as if they were written in the subclass (after translating type variables suitably if the base class is generic).
I think the re-check should be recursive, i.e. include bases of bases, etc. Of course, this could slow down the type-check, but not the runtime.
- A typechecker uses the normal MRO to find the signature of a method. For example, in this code:
class Base:
def method(self, x: int) -> None: ...
class AnotherBase:
def another_method(self, y: int) -> None: ...
class Derived(Implementation[Base], AnotherBase):
def method(self, x: str) -> None: ...
a typechecker infers that Derived().method has signature (str) -> None, and Derived().another_method has signature (int) -> None.
- Clearly,
Implementationcould be used only with a proper class.Implementation[Union[t1, t2]],Implementation[Callable[[t1, t2, ...], tr]], etc. riseTypeError.
If we agree on this, then I will submit PRs for typing.py, PEP484, and PEP483.
This has the problem that a type checker doesn't see inside the implementations of stubbed classes, and you could break some of the base class methods without a type checker having any way of detecting it.
Maybe implementation inheritance would only be allowed from within your own code somehow. For example, behavior would be undefined if you inherit the implementation of a stdlib or 3rd party library class, unless you include the implementation of the library module in the set of checked files. object would be fine as a base class, of course, but you probably shouldn't override any methods defined in object in an incompatible manner, since this can break a lot of things. This would only be enforced by static checkers, not at runtime.
Not sure if the latter would be too restrictive. To generalize a bit, it would okay if the implementation base class has library classes in the MRO, but methods defined in those classes can't be overridden arbitrarily.
@JukkaL This is a good point. We need to prohibit incompatibly overriding methods everywhere except for accessible in the set of checked files. I don't think it is too restrictive. Alternatively, we could allow this with a disclaimer that complete typecheking is not guaranteed in this case.
Yeah, I'm afraid the use case from https://github.com/python/mypy/issues/1237 is not so easily addressed by things that preserve soundness. But I don't have an alternate proposal. :-(
@gvanrossum Indeed, there is not much could be done. But I don't think this is bad. As with Any or # type: ignore, people who are going to use this feature know what they do. More important is that Implementation provides more precise typechecking, more "granularity", and probably a bit more safety than Any.
The QPixmap example shouldn't be a problem, since it seems to be for a library stub file. We don't verify that stubs are correct anyway, so a stub would be able to use implementation inheritance without restrictions but there are no guarantees that things are safe -- we expect the writer of the stub to reason about whether implementation inheritance is a reasonable thing to use.
For user code, I think that we can provide a reasonable level of safety if we enforce the restrictions that I mentioned.
I have a related proposal. What about having a magic class decorator (with no runtime or minimal effect) or a base class that declares a class as a "pure mixin"? A pure mixin could only be used for implementation inheritance, and it wouldn't be type checked by itself at all -- it would only be type checked as part of implementation inheritance. I think that this would allow type checkers to check some mixin use cases that are tricky right now with minimal code changes. The alternative would be to use ABCs to declare the things that the mixin expects to be defined elsewhere, but they are harder to use and arguably add boilerplate.
Example:
@puremixin
class MyMixin:
def do_stuff(self) -> int:
self.one_thing() # ok, even though not defined in this class
return self.second_thing(1) # also ok
class Thing(Implementation[MyMixin]):
def one_thing(self) -> None: pass # ok
def second_thing(self, x: int) -> str: # error: return type not compatible with do_stuff
return 'x'
But is that the way mixins are typically used? It would be good to look carefully at some examples of mixins (e.g. one of our internal code bases uses them frequently).
@JukkaL How about the following disclaimer to be added?
""" Note that implementation inheritance could be used with stubbed extension modules, but there are no guarantees of safety in this case (type checkers could not verify correctness of extension modules). The writer of the stub is expected to reason about whether implementation inheritance is appropriate.
In user code, implementation inheritance is only allowed with classes defined in the set of checked files. Built-in classes could be used as implementation base classes, but their methods could not be redefined in an incompatible way. """
I like you proposal of @puremixin decorator (maybe we could choose other name). It looks like it is independent of the initial proposal and could be easily added. I propose @gvanrossum to decide whether we need this.
@gvanrossum I've seen mixins like that (that use things that aren't defined anywhere in the mixin class). I'm not sure how common they are.
@ilevkivskyi Looks good. I'd do some small changes (feel free to edit at will):
Note that implementation inheritance could be used with stubbed extension modules, but there are no guarantees of safety in this case (type checkers could not verify correctness of extension modules). The writer of the stub is expected to reason about whether implementation inheritance is appropriate.
->
Note that implementation inheritance could be used with stubbed extension modules, but there are no guarantees of safety in this case (type checkers could not verify correctness of extension modules). As a special case, if a stub uses implementation inheritance within the stub, the writer of the stub is expected to reason about whether implementation inheritance is appropriate. Stubs don't contain implementations, so type checker can't verify the correctness of implementation inheritance in stubs.
Built-in classes could be used ...
->
Built-in and library classes could be used ...
@gvanrossum I've seen mixins like that (that use things that aren't defined anywhere in the mixin class). I'm not sure how common they are.
We're probably talking about the same mixin classes. :-) However, I'm not convinced that those would deserve being marked as Implementation[] -- the class that inherits from the mixin typically also inherits from some other class and the mixin should not conflict with any methods defined there (though it may override them). (OTOH maybe I didn't read the examples all the way through.)
I was writing a proposal and it may be a bit messy to post it now because I duplicated some stuff already said, but wrote some other things differently. Let me summarize different or subtle points that I think should be discussed:
- Using a generic-like syntax like Implementation[Foo] looks to similar to generics while what we have is somewhat different. I've been thinking on a functional syntax, i.e.:
class Base: ...
class Derived(Foo, implementation_only(Base), Bar): ...
so it doesn't make sense to pass it a Typevar or anything that has no supporting class (like unions or protocols). I think the different syntax helps.
- other names I considered for implementation_only() are reuse() or include() or insert() (the latter ones makes sense if you think of the mechanism it implements). "implementation" sounds a bit vague (even if it was my first proposal)
- probably the type checker should verify that implementation_only is used only in the context of a class definition in the base class list.
- the typechecker should ignore the Base methods if overriden in Derived, or in other parent (Foo, but not Bar in the example above)
- inherited bodes should be checked if available; if stubs only no checking its done (it's consistent with the rest of mypy that ignore bodies if stubs are present)
-
super()needs some special care. The simplest way to approach it is that a super() object does not have methods that have a signature coming from an implementation_only parent, otherwise you could get mistyped calls from the mro chain. If you're doing implementation only inheritance, following the mro makes no sense anyway. We should allow some way to get the parent methods, maybe through an explicit class reference (allowing some leeway on the self parameter when some method on Derived saysBase.method(self, ...)) - A class defined having all its parents as implementation_only, is interpreted by the typechecker as a direct descendent of object.
As a side note, using functional syntax would make it trivial to use @implementation_only as a class decorator too, which can be used for the mixin mechanism proposed by @JukkaL . However for that example I'd prefer the mixin to actually have the signatures of its "abstract methods"; if I want to implement a mixin one of the things I want from a static type system id to give me some hints of what these methods should take. For me, the decorator would only make subclasses of the mixin to behave like the rest of this proposal
Btw, for further reference there's a paper on the idea at http://smarteiffel.loria.fr/papers/insert2005.pdf
There are some things that can be added to this mechanism (I left these apart because they aren't part of the core proposal to keep the above KISS).
- Allow classes to be always implementation_only (like @JukkaL said):
@implementation_only
class SomeMixin:
def method(self, x: int) -> str: ...
def other_method(self, x: int) -> str: ...
class C(SomeMixin):
def method(self) -> None: ... # OK to change the signature
- Allow to selectively include or exclude parts of the parent class:
class C(
implementation_only(
SomeMixin, include=('method', 'attribute')
)
): ...
class D(
implementation_only(
SomeMixin, exclude=('other_method')
)
): ...
- Allow renames of methods
class E(
implementation_only(
SomeMixin, renames={'method': '_d_method'}
)
): ...
The exclusion/renaming mechanism are present in Eiffel, and the renaming allows to unambiguously resolve issues related to super() by given explicit different names to the overriden methods, although the mechanism has some complexities that might be problematic in python.
I don't think the first proposal should include any of these, but I wanted to mention these to also clarify that the functional syntax gives more flexibility for future growth of the idea.
@dmoisset Indeed, some things you propose have already been said. This is rather a good sign. Some comments:
- I don't think that we need functional syntax, square brackets are used not only for generics they are used almost everywhere in the type checking context.
- I like the name
Implementation. It reminds me that I am inheriting only the "implementation" of the class, not the "interface" (i.e. type contracts). - I am not sure that the restriction about
super()is really necessary.
The possible problem with super happens in the following snippet:
def implementation_only(C): return C
class Base:
def f(self, a: int) -> None:
print("Base")
print(a)
class Left(Base):
def f(self, a: int) -> None:
print("Left.f")
super().f(a)
class Right(implementation_only(Base)):
def f(self, a: str) -> None:
print("Right.f")
super().f(len(a))
class C(Left, implementation_only(Right)):
def f(self, a: int) -> None:
print("C.f")
super().f(a)
C().f(42)
This code fails with a TypeError if run with Python. but the rules defined above don't detect a problem:
-
BaseandLeftare correctly typed -
Rightredefines the signature off(which is ok, given that it's using implementation inheritance). -
Right.fcallssuper.fusing an int argument (which looks ok, given that the inheritedfaccepts an int) -
Credefines the signature ofRight.f(which is ok, given that it is using implementation inheritance). It keeps the signature ofLeft.fso no problem there -
C.fcallssuper.fwith an int. The type checker is only looking at theLeft.fsignature, so this is considered valid.
But even if everything looks ok by itself, the program has a type error, so the rules are unsound. So some of the rules above should actually be rejected in this scenario. 1 seems to be an ok rule (valid with the current semantics), and 2 is our proposal so let's assume it as valid (if we get to a contradiction the idea itself is unsound). So 3, 4, or 5 should be wrong, and we should choose (at least) one, under some conditions (which should include this scenario). My proposal was "3 is wrong. You can't call super from an implementation_only redefinition of a method". This actually will make the call that produces the TypeError in runtime to be invalid.
Other options are making this scenario impossible forbidding the redefinition in some multiple inheritance cases (4 would be invalid), but it's probably harder to explain to users
Thank you for elaborating, I understand your point. However, it looks like a more sophisticated typechecker could catch the TypeError. By re-type-checking bodies of methods in C parent classes, a typechecker could detect that the super().f(a) call in Left has incompatible signature, since super() for Left is now Right, and f in Right requires a str.
Prohibiting 3 is an absolutely valid option (although it feels a bit false positive to me), also prohibiting 4 is a valid option. What I propose is to leave this choice (sophisticated recursive re-checking vs prohibitng 3 vs prohibiting 4) up to the implementer of a particular typechecker.
Restrictions mentioned by Jukka are necessary, because it is extremely difficult (probably impossible in principle) to check types inside extension modules, so that we: a) "forgive" typecheckers all uncaught TypeErrors related to this; b) propose some restrictions for user code that nevertheless allow to maintain safety. At the same time, typecheckers should be able to catch the error like you show in either way they choose (i.e. if a typechecker is not able, then this could be considered a bug in that checker).
I have to say that I am uncomfortable with solutions (to any problem, not just this one) that say "just re-type-check such-and-such code with this new information". This could potentially slow things down tremendously (depending on how often new information is discovered and acted upon).
It also makes things hard to explain to users -- e.g. how to explain that their code that was previously marked as correct is now deemed incorrect, due to something that someone else added somewhere else. Why not place the blame on the new subclass that apparently violates the assumptions of the base class, for example?
I also worry that such solutions violate some fundamental nature of type annotations. When mypy sees "def foo(arg: int) -> int: ..." it type-checks the body of foo() and then it can forget about it -- everything it needs to know about foo() is now summarized in the function signature. Likewise for classes. And for modules. A solution that requires mypy to go back essentially introduces a circular dependency between (in this case) base class and superclass, or maybe (in other examples) caller and callee, or importer and importee.
Those circular dependencies worry me especially in the context of the work we're doing on incremental checking in mypy -- it makes the assumption that if a module hasn't changed, and its dependencies (taken transitively) haven't changed either, then the module doesn't need to be re-checked on a future run, and the information collected in its symbol table (about global variables, functions, classes, methods and signatures) is sufficient to analyze any new code that happens to import it. I'd like not to see this assumption violated.
Of course it's common that circular dependencies occur at all levels -- import cycles, mutually recursive functions, what have you. And mypy deals with them (and we're improving it for those edge cases where it fails to deal with them). But I think I want to avoid introducing extra cycles in relationships that users see as one-directional.
The point about incremental checking is a good one. General performance could also be an important concern if people would start using this more widely. My expectation would be that this would mostly be used in fairly simple cases, but of course any feature given to users tends to be used in unforeseen ways, and once the genie is out of the bottle it's difficult to put it back there. There could be also other interactions with particular language features that we haven't though of yet.
Due to all the concerns and tricky interactions that we've uncovered, I'd like to see a prototype implementation of this before deciding whether to add this to the PEP -- I have a feeling that we don't yet understand this well enough.
After thinking more it looks like there are three important "use cases" for this feature:
- The original request originates from typing extension modules;
- Another use case is inheriting from library classes, where a user has some already working code, that redefines methods in a "theoretically" incompatible way;
- A use case proposed by @JukkaL: pure mixin as a class that is not typechecked at all when defined, but only when inheriting from it.
What I see as important now is that first two use cases actually do not require recheck of base classes (and it is even impossible for built-in types, extension modules and library stubs), only the third one requires it. Let us focus on first two cases here and discuss re-typechecking and related issues in #246.
A reworked proposal that covers two first cases is: Implementation is just a way to say "I know what I do" that is "softer" than Any. More details:
- At runtime
Implementation[MyClass] is MyClass. All normal mechanisms work at runtime:
class Base:
def __init__(self, number: int) -> None:
self.number = number
class AnotherBase:
...
class Derived(Implementation[Base], AnotherBase):
def __init__(self) -> None:
super().__init__(42) # This works since Implementation[Base] is just Base
- The derived class is not considered a subtype of implementation base classes. An explicit cast might be used under user's responsibility. Examples:
from typing import cast
def func(x: Base) -> None:
...
def another_func(x: AnotherBase) -> None:
...
func(Base(42)) # this is OK
another_func(Derived()) # this is also OK
func(Derived()) # Marked as error by a typechecker
func(cast(Base, Derived())) # Passes typecheck
- Methods of an implementation base class could be redefined in an incompatible way:
class Base:
def method(self, x: int) -> None:
...
class AnotherBase:
def another_method(self, y: int) -> None:
...
class Derived(Implementation[Base], AnotherBase):
def another_method(self, y: str) -> None: ... # This is prohibited by a typechecker
def method(self, x: str) -> None: ... # Although, this is OK
No re-typecheck of base classes is performed, it is the responsibility of a user to ensure that the new signature is safe.
- A typechecker uses the normal MRO to find the signatures of methods and uses them to typecheck their use in derived class and elsewhere as usual. For example:
class Base:
def method(self, x: int) -> None: ...
class AnotherBase:
def another_method(self, y: int) -> None: ...
class Derived(Implementation[Base], AnotherBase):
def method(self, x: str) -> None: ...
Derived().method('abc') # OK
Derived().another_method('abc') # Fails, incorrect type of argument
Derived().third_method('abc') # Fails, no such method
-
Implementationcould be used only with a proper class.Implementation[Union[t1, t2]],Implementation[Callable[[t1, t2, ...], tr]], etc. raiseTypeError. -
People who will use this feature probably want some speed, therefore implementation of
Implementationshould be fast. Probably, it should not be a class itself, just an object whose class defines__getitem__that returns its argument.
@gvanrossum @dmoisset what do you think about this?
It still sounds like too big a hammer. In https://github.com/python/mypy/issues/1237#issuecomment-188485241 @philthompson10 wrote:
Is there a case for treating stub files differently, maybe by providing something like @overload that tells mypy that Bar.factory() is not an override of Foo.factory()?
From this it seems pretty clear that he doesn't want to completely sever the ties between Foo and Bar, only between the two unrelated factory() methods. (Though perhaps this is a job for # type: overload?)
@gvanrossum I think it was more or less established that the initial example in #1237 (the Foo.factory return type definition) was unsound, and that case can already be solved with a #type: ignore at the redefinition point, without affecting the relationship between classes.
The other scenario presented later (The one about QBitmap) may be more interesting (I'm assuming those classes can be used differently.
To move away from examples that are not my own, I also found a similar one at https://github.com/django/django/blob/stable/1.10.x/django/utils/datastructures.py#L48 ; there you have a MultiValueDict class that inherits dict because the implementation of dict is good for it, but it has a slightly different API (actually, it has the API of a Dict[K, V] extended, but the implementation of a Dict[K, List[V]]). It's ok because it's not intended to replace a dict, it's just a new data structure to be used separately.
In this situation, I think a proposal like the last one from @ilevkivskyi 's would work. There is some potential for unsoundness (essentially every reference to "self" in the inherited class can be subject to a broken call), but the alternatives here are using Any or # type: ignore which can also lead to runtime errors. The point here is to be able to explicitly say "this inherits from dict, but it's not something you can use in place of a dict", and have that checked (and allow me to break the dict contract in a class that is not really a dict).