Unable to specify a default value for a generic parameter

[rowillia]

Simplified Example:

from typing import TypeVar

_T = TypeVar('_T')
def foo(a: _T = 42) -> _T:  # E: Incompatible types in assignment (expression has type "int", variable has type "_T")
    return a

Real World example is something closer to:

_T = TypeVar('_T')
def noop_parser(x: str) -> str:
    return x

def foo(value: str, parser: Callable[[str], _T] = noop_parser) -> _T:
    return parser(value)

[gvanrossum]

I think the problem here is that in general, e.g. if there are other parameters also using _T in their type, the default value won't work. Since this is the only parameter, you could say that this is overly restrictive, and if there are no parameters, the return type should just be determined by that default value (in your simplified example, you'd want it to return int). But I'm not so keen on doing that, since it breaks as soon as another generic parameter is used.

To make this work, you can use @overload, e.g.

@overload
def foo() -> int: ...
@overload
def foo(a: _T) -> _T: ...
def foo(a = 42):  # Implementation, assuming this isn't a stub
    return a

That's what we do in a few places in typeshed too (without the implementation though), e.g. check out Mapping.get() in typing.pyi.

[OddBloke]

I'm running in to a similar issue when trying to iron out a bug in the configparser stubs. Reducing RawConfigParser to the relevant parts gives:

_section = Mapping[str, str]

class RawConfigParser:
    def __init__(self, dict_type: Mapping[str, str] = ...) -> None: ...

    def defaults(self) -> _section: ...

This is incorrect at the moment, because dict_type should (as the name indicates) be a type. Furthermore, defaults() returns an instance of dict_type, so this seems like a good application of generics. The following works as expected when a dict_type is passed (i.e. reveal_type(RawConfigParser(dict_type=dict).default()) gives builtins.dict*[Any, Any]):

_T1 = TypeVar('_T1', bound=Mapping)

class RawConfigParser(Generic[_T1]):
    def __init__(self, dict_type: Type[_T1] = ...) -> None: ...

    def defaults(self) -> _T1: ...

but when you don't pass a dict_type, mypy can't infer what type it should be using and asks for a type annotation where it is being used. This is less than ideal for the default instantiation of the class (which will be what the overwhelming majority of people are using). So I tried adding the appropriate default to the definition:

_T1 = TypeVar('_T1', bound=Mapping)

class RawConfigParser(_parser, Generic[_T1]):
    def __init__(self, dict_type: Type[_T1] = OrderedDict) -> None: ...

    def defaults(self) -> _T1: ...

but then I hit the error that OP was seeing:

stdlib/3/configparser.pyi:51: error: Incompatible default for argument "dict_type" (default has type Type[OrderedDict[Any, Any]], argument has type Type[_T1])

This feels like something that isn't currently supported, but I'm not sure if I'm missing a way that I could do this...

[ilevkivskyi]

@OddBloke
This default is, strictly speaking, not safe. This would require a lower bound for _T1 in order to be safe.

[OddBloke]

@ilevkivskyi I don't follow, I'm afraid; could you expand on that a little, please?

[ilevkivskyi]

Consider this code:

class UserMapping(Mapping):
    ...
RawConfigParser[UserMapping]()

In this case OrderedDict will be "assigned" to dict_type, which will be Type[UserMapping], which is not a supertype of Type[OrderedDict]. In order for the default value to be safe it is necessary to have some guarantee that _T1 will be always a supertype of OrderedDict (this is called a lower bound), but this is not supported, only upper bounds are allowed for type variables.

[OddBloke]

@ilevkivskyi Thanks, that makes sense. Do you think what I'm trying to do is unrepresentable as things stand?

[ilevkivskyi]

I didn't think enough about this, but my general rule is not to be "obsessed" with precise types, sometimes Any is OK. If a certain problem appears repeatedly, then maybe we need to improve something (also now we have plugin system). In this particular case my naive guess would be to play with overloads, as Guido suggested (note that __init__ can also be overloaded).

[nematsakis]

Allowing default values for generic parameters might also enable using type constraints to ensure that generics are only used in specified ways. This is a bit esoteric, but here is an example of how you might use the same implementation for a dict and a set while requiring that consumers of the dict type always provide a value to insert() while consumers of the set type never do.

class Never(Enum):
    """An uninhabited type"""

class Only(Enum):
    """A type with one member (kind of like NoneType)"""
    ONE = 1

NOT_PASSED = Only(1)

class MyGenericMap(Generic[K, V]):

    def insert(key, value=cast(Never, NOT_PASSED)):
        # type: (K, V) -> None
        ...
       if value is NOT_PASSED:
           ....

class MySet(MyGenericMap[K, Never]):
    pass

class MyDict(MyGenericMap[K, V]):
    pass

my_set = MySet[str]()
my_set.insert('hi')           # this is fine, default value matches concrete type
my_set.insert('hi', 'hello')  # type error, because 'hello' is not type never

my_dict = MyDict[str, int]()
my_dict.insert('hi')          # type error because value is type str, not Never
my_dict.insert('hi', 22)      # this is fine, because the default value is not used

[gvanrossum]

Huh, this just came up in our code.

[ilevkivskyi]

This request already appeared five times, so I am raising priority to high.

I think my preferred solution for this would be to support lower bounds for type variables. It is not too hard to implement (but still a large addition), and at the same time it may provide more expressiveness in other situations.

[hynekcer]

A second part for the workaround with overload should be noted: The header of implementation of the overloaded function must be without a generic annotation for that variable, but it is not acceptable for more complex functions where the body should be checked. Only the binding between input and output types is checked by overloading. The most precise solution for the body seems to encapsulate it by a new function with the same types, but without a default.

@overload
def foo() -> int: ...
@overload
def foo(a: _T) -> _T: ...

def foo(a = 42):  # unchecked implementation
    return foo_internal(a)

def foo_internal(a: _T) -> _T:
    # a checked complicated body moved here
    return a

Almost everything could be checked even in more complicated cases, but the number of necessary overloaded declarations could rise exponentially by 2 ** number_of generic_vars_with_defaults.

EDIT

[hynekcer]

Another solution is to use a broad static type for that parameter and immediately assign it in the body to the original exact generic type. It is a preferable solution for a class overload.
An example is a database cursor with a generic row type:

_T = TypeVar('_T', list, dict)

class Cursor(Generic[_T]):
    @overload
    def __init__(self, connection: Connection) -> None: ...
    @overload
    def __init__(self, connection: Connection, row_type: Type[_T]) -> None: ...
    def __init__(self, connection: Connection, row_type: Type=list) -> None:
        self.row_type: Type[_T] = row_type  # this annotation is important
        ...
    def fetchone(self) -> Optional[_T]: ...
    def fetchall(self) -> List[_T]: ...   # more methods depend on _T type

cursor = Cursor(connection, dict)  # cursor.execute(...)
reveal_type(cursor.fetchone())     # dict

[andersk]

I think the problem here is that in general, e.g. if there are other parameters also using _T in their type, the default value won't work.

Wait. Why is that a problem? If default value doesn’t work for some calls, that should be a type error at the call site.

_T = TypeVar('_T')
def foo(a: List[_T] = [42], b: List[_T] = [42]) -> List[_T]:
    return a + b

foo()  # This should work.
foo(a=[17])  # This should work.
foo(b=[17])  # This should work.
foo(a=["foo"])  # This should be a type error (_T cannot be both str and int).
foo(b=["foo"])  # This should be a type error (_T cannot be both int and str).
foo(a=["foo"], b=["foo"])  # This should work.

This default is, strictly speaking, not safe. This would require a lower bound for _T1 in order to be safe.

Only if we retain the requirement that the default value always works, which is not what’s being requested here. The desired result is:

_T1 = TypeVar('_T1', bound=Mapping)
class RawConfigParser(_parser, Generic[_T1]):
    def __init__(self, dict_type: Type[_T1] = OrderedDict) -> None: ...

    def defaults(self) -> _T1: ...

class UserMapping(Mapping):
    ...

RawConfigParser[OrderedDict]()  # This should work.
RawConfigParser[UserMapping]()  # This should be a type error (_T1 cannot be both UserMapping and OrderedDict).
RawConfigParser[UserMapping](UserMapping)  # This should work.

[radeksz]

I ran into this trying to write

_T = TypeVar('_T')
def str2int(s:str, default:_T = None) -> Union[int, _T]: # error: Incompatible default for argument "default" (default has type "None", argument has type "_T")
    try:
        return int(s)
    except ValueError:
        return default

(after figuring out that I need --no-implicit-optional mypy parameter as well). @overload is not ideal because then the body of the function is no longer type-checked.

@overload
def str2int(s:str) -> Optional[int]: ...
@overload
def str2int(s:str, default:_T) -> Union[int, _T]: ... 

def str2int(s, default = None): 
    the_body_is_no_longer_type_checked

I think this raises the bar for type-checking newbies.

[markedwards]

To @andersk's point, I also don't see how PEP 696 makes any difference here. The TypeVar default does not address the core issue, which is that the default value might conflict with the type variable.

For what its worth, Typescript also does not support this pattern:

And adding a type variable default in Typescript does not "fix" the issue:

That's not to say that Python typing is for any reason bound by what Typescript can do, just that this limitation is not a peculiarity of Python or mypy.

It seems to me that a final decision should be made here and this should be documented so developers don't need to find this thread to sort it out.

[NeilGirdhar]

To @andersk's point, I also don't see how PEP 696 makes any difference here.

Sorry, I mistakenly brought up PEP 696. I hadn't looked at this issue carefully since I subscribed to it, and forgot what it was actually about.

[JukkaL]

I'm currently doing some experimentation in the hope of eventually addressing this.

[JukkaL]

I haven't made much progress with an implementation yet, but here are some notes based on what I've learned so far (much of this is also covered above in the discussion, but I want so summarize everything in one comment):

• Storing the types of default values in a callable type and using these to generate additional constraints when default arguments aren't provided at a call site seems like the most promising approach. We'd only store the default value types for generic functions that have default arguments to minimize the impact to memory use and performance.
• We can't infer the types of default values during semantic analysis. This means that we can't always produce the signature of a function during semantic analysis any more. A possible workaround is to infer the signature of a generic callable that accepts defaults lazily during type checking, similar to what we do with decorated functions. We probably don't want to do this for all functions, since it could be quite a disruptive change.
• We need to make the check for valid default argument values in a function definition more lenient.
• For arguments such as x: T | None = None, we wouldn't generate a constraint since the value of the default value is also covered by the argument type. These cases may have additional trickiness as well.

[ilevkivskyi]

Yeah, this is essentially a plan I had in mind a while ago. But after all it seemed to me this would cause a visible performance impact even with all the possible optimizations. OTOH it may be fine with all the recent performance improvements. Btw few more points to keep in mind:

• We may need to special-case typeshed stubs (and maybe protocols) in cases like def foo(x: T = ...) -> T: ... (with literal ... as default). Also this change will mean providing actual default will change the type in stubs, i.e. def foo(x: T = ...) -> T: ... and def foo(x: T = 0) -> T: ... would be two different types.
• All the type ops for callables (and other types, e.g. overloads) like is_subtype, meet, join, etc., will need to be reconsidered to account for default types.
• There may be some trickiness if default values are lambdas, since there is a very old bug when applying generic function to a generic function, it may cause troubles here.

[couling]

A second part for the workaround with overload should be noted: The header of implementation of the overloaded function must be without a generic annotation for that variable, but it is not acceptable for more complex functions where the body should be checked. Only the binding between input and output types is checked by overloading.

May I ask if there's a little more detail on this design decision? Why are funciton implementations not checked for overloads?

Searches on the topic are giving me blanks. I'm curious why mypy would not check a defanition for each overload variant. It would only need to revist one function per overload so I can't see the cost would be that high, nor the implementation that complex. I wonder what I'm missing?