jellezijlstra / stubcheck Goto Github PK

These:

+selectors: 'DevpollSelector' is in stub but is not defined at runtime
 +selectors: 'EpollSelector' is in stub but is not defined at runtime
 +selectors: 'PollSelector' is in stub but is not defined at runtime

happen because the selectors only exist on OSs with the underlying syscalls available.

stubcheck should either (1) implement a clever way to detect that these are OS-specific or (2) provide a way to whitelist specific names. The only way I can think of to do (1) is to do source code analysis on the module and find that DevpollSelector is defined inside an if block, but that seems hacky and won't work for C modules anyway, so I'll probably go with (2).

Proposed format: We add a file tests/stubcheck_blacklist.txt to typeshed. Every line in this file is either a comment starting with #, a blank line, or of the form selectors.DevpollSelector. Such a line tells stubcheck to suppress all errors related to selectors.DevpollSelector. To suppress all errors in the selectors module, we could accept a glob like selectors.*. Alternatively, it could accept file names, so stdlib/3.4/selectors.pyi will suppress all errors in that file. But that would block stubcheck from finding other errors: for example, if DevpollSelector does exist, it should still be able to check that its methods have the right signature.

Currently stubcheck reports

+xml: 'dom' is in __all__ but not in stub
 +xml: 'etree' is in __all__ but not in stub
 +xml: 'parsers' is in __all__ but not in stub
 +xml: 'sax' is in __all__ but not in stub

At runtime, these are in __all__ but they're not directly imported in xml/__init__.py, so the stub is correct. stubcheck needs to detect this situation. It also affects the email package.