-
-
Notifications
You must be signed in to change notification settings - Fork 3.1k
Description
Currently, mypy ensures monotnonically increasing order when testing overloads against overloads:
Lines 1003 to 1009 in 9f157d9
| # Order matters: we need to make sure that the index of | |
| # this item is at least the index of the previous one. | |
| if subtype_match and previous_match_left_index <= left_index: | |
| previous_match_left_index = left_index | |
| found_match = True | |
| matched_overloads.add(left_index) | |
| break |
However, I believe (1) this rule does more harm than good in many cases when it is correct, and (2) there are many cases where it is incorrect. Therefore, I propose to move this check behind a new strictness flag and disable it by default. Later, one can see if the check can be refined to deal with the cases listed below.
Rationale
In python/typing#2021 I showed that the current typing spec is too strict with respect to subtyping overloads.
When this check is incorrect
The check is incorrect when overloads are commutative, i.e. changing the overload order describes the same callable type. This is the case when:
-
Argument signatures are mutually exclusive mypy-playground
-
Argument types are mutually exclusive (e.g. via
@disjoint_bases) mypy-playground -
Argument types overlap, but return types are identical. mypy-playground
When this check is technically correct, but hurts in practice
As @disjoint_bases is relatively new, many libraries likely do not use it. Moreover, Intersection types are still not supported. Thus, many classes that are compatible at runtime cannot be put under the umbrella of a shared Protocol because inadvertently they will use different overload order at one point or another.