typeshed

About

Typeshed models function types for the Python standard library and Python builtins, as well as third party packages.

This data can e.g. be used for static analysis, type checking or type inference.

Format

Each Python module is represented by a .pyi "stub". This is a normal Python file (i.e., it can be interpreted by Python 3), except all the methods are empty. Python function annotations (PEP 3107) are used to describe the types the function has. See PEP 484 for the exact syntax of the stub files.

Syntax example

The below is an excerpt from the types for the datetime module.

MAXYEAR = ...  # type: int
MINYEAR = ...  # type: int
__doc__ = ...  # type: str
__file__ = ...  # type: str
__name__ = ...  # type: str
__package__ = ...  # type: None

class date(object):
    def __init__(self, year: int, month: int, day: int) -> None: ...
    @classmethod
    def fromtimestamp(cls, timestamp: int or float) -> date: ...
    @classmethod
    def fromordinal(cls, ordinal: int) -> date: ...
    @classmethod
    def today(self) -> date: ...
    def ctime(self) -> str: ...
    def weekday(self) -> int: ...

Directory structure

Builtins vs stdlib

Python ships with a set of built-in modules, i.e., modules that are baked into the Python executable. For a specific Python build, you can use sys.builtin_module_names to query which modules are built in. Also, you can determine whether a module is a built-in module by doing import module; import.__file__. If __file__ exists, the module is not built in. Typeshed stores built-in modules in the "builtins/" directory. Examples for built-in modules: sys, array, math, signal.

There are other modules that ship with Python, but are not linked into the Python binary. E.g. os.py, glob.py, zipfile.py. (But also some C extensions like datetime) These modules are stored in the stdlib/ directory.

Note that built-in modules have higher precedence in the import path than stdlib modules. The former are implicitly prepended to the start of your PYTHONPATH, whereas the latter are implicitly appended to it.

Addendum: Some Linux distributions ship Python built so that non-essential builtins, like datetime, are instead a shared library in lib-dynload/. In typeshed, we will treat these as builtins as well, because they come with Python and are implemented in C.

third_party

Modules that are not shipped with Python but have a type description in Python go into third_party. Since these modules can behave differently for different versions of Python, third_party has version subdirectories, just like stdlib and builtins.

We're welcoming contributions (pull requests) for type definitions of third party packages.

Version directories

We store stubs for both Python 2 as well as Python 3. We also distinguish between minor versions (E.g. 3.2 <-> 3.3). To accomplish not having to duplicate modules that are the same between all minor versions, we have e.g. a top-level directory 3/ that contains all the stubs for Python 3. More specialized stubs go into e.g. 3.3/ and supersede the more generic stubs in 3/. (And, if needed, a directory 3.3.1/ would be able to supersede stubs in 3.3/). Modules that are the same under both Python 2 and Python 3 go into 2and3/.

Combining multiple versions in a single file

According to PEP 484, type checkers are expected to understand simple version and platform checks. So the following syntax is legal in a pyi:

if sys.version_info[0] >= 3:
    # Python 3 specific definitions
else:
    # Python 2 specific definitions

This can be used for modules in 2and3/ that only have minor changes between Python 2 and Python 3. If the difference between versions is more drastic, it can make more sense to have seperate files in 2.x/ and 3.x/.

Directory structure

Directory	Contents
builtins/2and3/	Builtin stubs for Python 2 and Python 3
builtins/2/	Builtin stubs for Python 2
...	...
builtins/2.7/	Builtin stubs for Python 2.7
builtins/3/	Builtin stubs for Python 3
...	...
builtins/3.3/	Builtin stubs for Python 3.3 (replacing generic stubs in 3/)
stdlib/2and3/	Standard library stubs for Python 2 and Python 3
stdlib/2.7/	Standard library stubs for Python 2.7
...	...
stdlib/2.7.6/	Standard library stubs specialized for Python 2.7.6
...	...
third_party/2and3/	Third party modules for Python 2 and 3
...	...

Scripts for generating stubs

Scripts for generating stubs should not go into typeshed itself.

Ben Longbons is maintaining a repository with code for auto-generating stubs or creating prototypes of stubs. It's at o11c/stubtool.