何时在 Python 类型提示中使用 IO[str]/IO[bytes] 和 TextIO/BinaryIO？

Question

From the documentation, it says that:

Generic type IO[AnyStr] and its subclasses TextIO(IO[str]) and BinaryIO(IO[bytes]) represent the types of I/O streams such as returned by open() .

— Python Docs:typing.IO

The docs did not specify when BinaryIO / TextIO shall be used over their counterparts IO[str] and IO[bytes] .

Through a simple inspection of the Python Typeshed source, only 30 hits found when searching for BinaryIO , and 109 hits for IO[bytes] .

I was trying to switch to BinaryIO from IO[bytes] for better compatibility with sphinx-autodoc-typehints , but the switch-over has broken many type checks as methods like tempfile.NamedTemporaryFile is typed as IO[bytes] instead of the other.

Design-wise speaking, what are the correct situations to use each type of these IO type hints?

Answer 1

BinaryIO and TextIO directly subclass IO[bytes] and IO[str] respectively, and add on a few extra methods -- see the definitions in typeshed for the specifics.

So if you need these extra methods, use BinaryIO/TextIO . Otherwise, it's probably best to use IO[...] for maximum flexibility. For example, if you annotate a method as accepting an IO[str] , it's a bit easier for the end-user to provide an instance of that object.

Though all this being said, the IO classes in general are kind of messy at present: they define a lot of methods that not all functions will actually need. So, the typeshed maintainers are actually considering breaking up the IO class into smaller Protocols . You could perhaps do the same, if you're so inclined. This approach is mostly useful if you want to define your own IO-like classes but don't want the burden of implementing the full typing.IO[...] API -- or if you're using some class that's almost IO-like, but not quite.

All this being said, all three approaches -- using BinaryIO/TextIO , IO[...] , or defining more compact custom Protocols -- are perfectly valid. If the sphinx extension doesn't seem to be able to handle one particular approach for some reason, that's probably a bug on their end.