No `TypeVar` variable inference in ternary operators

[sasanjac]

I have searched for a similar problem to reproduce but I haven't found anything, so:

I using a self-implemented monad which roughly looks as follows:

class Ok[T, E = None]:

    def __init__(self, value: T) -> None:
        self._value = value

class Err[E, T = None]: 

    def __init__(self, value: E) -> None:
        self._value = value

type Result[T, E] = Ok[T, E] | Err[E, T]

Now, inference of the second TypeVar variable works as intended, apart from the ternary operator.

Actual Behavior

class[U] Bar:
    def foo(data: U, cond: bool) -> Result[U, str]:
        return Ok(data) if cond else Err("Error")

results in

Incompatible return value type (got "Ok[U, str] | Err[str, None]", expected "Ok[U, str] | Err[str, U]")

Expected Behavior

def foo[U](data: U, cond: bool) -> Result[U, str]:
    if cond:
         return Ok(data)
     else
         return Err("Error")

and

def foo[U](data: U, cond: bool) -> Result[U, str]:
    if cond:
         return Ok(data)
    
     return Err("Error")

typecheck just fine.

Your Environment

• Mypy version used:
  mypy 1.15.0 (compiled: yes)
• Mypy command-line flags:
  none
• Mypy configuration options from mypy.ini (and other config files):

  [tool.mypy]
    check_untyped_defs      = true
    disallow_any_unimported = true
    disallow_untyped_defs   = true
    follow_imports          = "normal"
    mypy_path               = "src"
    namespace_packages      = true
    no_implicit_optional    = true
    show_error_codes        = true
    strict_optional         = true
    warn_no_return          = true
    warn_redundant_casts    = true
    warn_return_any         = true
    warn_unused_ignores     = true

• Python version used:
  Python 3.13.2

[terencehonles]

I don't think this is an issue if the ternary expression. It looks like it's an issue how you define Err. In your type expression you default T = None and it's never used so there's no way mypy can infer something other than None which matches what your error is suggesting Ok[U, str] | Err[str, None].

If you rewrite your generic functions as class methods do they still work? (Just wondering if this is because one is using function generics rather than class generics)

For your two base types you never use one of the type parameters. Do you need the second type? So something like:

class Ok[T]:
    def __init__(self, value: T) -> None:
        self._value = value

class Err[E]: 
    def __init__(self, value: E) -> None:
        self._value = value

type Result[T, E] = Ok[T] | Err[E]

I would think this probably would work for you since if you drop the second parameters from the error you're seeing it would be as expected:

Incompatible return value type (got "Ok[U] | Err[str]", expected "Ok[U] | Err[str]")

[sasanjac]

Unfortunately, this approach doesn't work as well since there are methods that may convert the type of the inner value, e.g.:

def map[U, E](self, op: t.Callable[[T], U]) -> Result[U, E]:
    return Ok(op(self._value))

When only using one type parameter, the other parameter has to be provided using function generics which mypy can also not infer.

The following example would not work with your proposed implementation:

def bar(*, x: int, y: str) -> str:
    return str(x) + y


def foo(x: Result[int, str]) -> Result[str, str]:
    return x.map(lambda x: bar(x=x, y="foo"))

mypy infers map as:
(method) map: ((op: ((int) -> U@map)) -> Result[U@map, E@map]) | ((op: ((T@map) -> U@map)) -> Result[U@map, str])
and complains:
Argument "x" to "bar" has incompatible type "T"; expected "int"Mypy[arg-type](https://mypy.readthedocs.io/en/latest/_refs.html#code-arg-type)

With my implementation, map is (method) map: ((op: ((int) -> U@map)) -> Result[U@map, str])

What I find weird, is that inference using the ternary operator does not work, however when writing the verbose if-else, it works.

[terencehonles]

I see, but for your map case should you actually be writing that as?

def map[U](self, op: t.Callable[[T], U]) -> Ok[U]:
    return Ok(op(self._value))

I assume you could come up with more and more complex examples and you're likely getting out of mypy's coverage.

I believe the reason you're seeing different behavior between the ternary and the if/else expression is the ternary will be generating a type union that the constraint solver needs to match, whereas the if/else case is a simpler expression that it's more forgiving about.

[sasanjac]

This implementation actually gets out of mypy's coverage pretty quickly:

class State:
    def foo(self) -> None:
        pass

def bar() -> Result[State, str]:
     return Ok(State())

bar().map(lamba state: state.foo())

mypy complains:
"T" has no attribute "foo"

[sterliakov]

Mypy struggles with certain inference corner cases where the variables are indeterminate (here it's missing some connection between two params, it's clearly a bug). Only using the type params you actually need can greatly improve mypy life.

Here's a sample with redundant variables removed as @terencehonles suggested above - it does not issue any errors like you show. You probably tried to add another (now completely indeterminate) type variable T to Err.map? Here's what you can do instead:

from typing import Any, Callable, Never

class Ok[T]:
    def __init__(self, value: T) -> None:
        self._value = value

    def map[U](self, op: Callable[[T], U]) -> Ok[U]:
        return Ok(op(self._value))

    def map_err(self, op: Callable[[Any], object]) -> Ok[T]:
        return self
        
    def unwrap_or(self, default: object) -> T:
        return self._value

class Err[E]: 
    def __init__(self, value: E) -> None:
        self._value = value

    def map(self, op: Callable[[Any], object]) -> Err[E]:
        return self
        
    def map_err[U](self, op: Callable[[E], U]) -> Err[U]:
        return Err(op(self._value))

    def unwrap_or[U](self, default: U) -> U:
        return default

type Result[T, E] = Ok[T] | Err[E]

class State:
    def foo(self) -> str:
        return "foo"

def bar() -> Result[State, str]:
     return Ok(State())

reveal_type(
    bar().map(lambda state: state.foo()).map_err(str.lower).unwrap_or(0)
)

https://mypy-play.net/?mypy=master&python=3.12&flags=strict&gist=945eca5d6c87384bfd19b3f6d24a985e

[randolf-scholz]

Both #19562 and #19563 fix this.