Infer `Final` from all-caps-named variables

[m-aciek]

Feature

Constants are usually defined on a module level and written in all capital letters with underscores separating words. Examples include MAX_OVERFLOW and TOTAL. (ref. PEP 8)

It's de-facto standard in the industry, therefore we could infer Final from variables written in all capital letters.

Pitch

I think that this behaviour would match natural programmers expectations. Currently the need of using all-caps and a Final type annotation seems not being minimal.

MAX_SIZE = 9000
MAX_SIZE += 1  # Error reported by type checker

class Connection:
    TIMEOUT = 10

class FastConnector(Connection):
    TIMEOUT = 1  # Error reported by type checker

[jorenham]

I thought this was already the case. So if it isn't, then it probably should.

[sterliakov]

I'm -1 on this (but my voice shouldn't have a lot of weight). Python type system is designed to be very explicit, and that's a good thing - we do not "infer" any non-trivial type-level properties of objects not caused by explicit type hints.

AFAIC the only such (tiny) feature where var name affects typechecking results is defining multiple adjacent decorated functions named _ like this. I don't remember any other such precedent in mypy code.

First obvious problem is conditional assignment to otherwise constant variables ("eventually final"):

FOO = 1
if ...:
    FOO = 2  # Foo is not declared Final but would become implicitly final, breaking this

I do consider FOO a module-level constant in this case (semantically it's only set once, even if it takes multiple steps), but it cannot be Final.

Doing this would also reduce our typing spec conformance.

Finally, this would mean that a package that uses SCREAMING_SNAKE_CASE for some non-Final module attributes will have no way to make mypy happy without breaking its runtime public interface. For example, if a package supports configuration by assignments to its config and uses SCREAMING_SNAKE_CASE for config attributes, there will be no way to tell mypy "this should not be treated as final" without renaming all such attrs (semver-major breaking change). Yes, that's not the best possible style, but I've certainly seen that done in real packages.