​​Modular Typeshed

​​Agenda

  • ​​Vision for typeshed
  • ​​Which modules should be in typeshed?
  • ​​Mass generating stubs using stubgen
  • ​​Scaling typeshed to thousands of packages

​​Monolithic Typeshed

  • ​​Entire typeshed is shipped as a single package
  • ​​A buggy stub can be a big problem for users
  • ​​Slow iteration speed on fixes
  • ​​Can’t pin to particular version of stub

​​Monolithic Typeshed

  • ​​Can’t have stubs for multiple versions of a package
  • ​​Users of old package versions are out of luck
  • ​​Users may be stuck with old versions of type checkers due to typeshed
  • ​​Problems will only get worse if we grow typeshed aggressively

​​Vision for Typeshed

  • ​​Modular typeshed: separate PyPI package for each stub package (PEP 561)
  • ​​Typical Python user can find stubs for every package they use
  • ​​Stubs for thousands of popular 3rd party packages available
  • ​​Fast iteration speed: new stub package ~immediately available after PR has been merged
  • ​​Bugs in stubs are less bad
  • ​​Move fast & fix things

​​Challenge 1: Which Packages to Include

  • ​​Thousands of PyPI packages 
  • ​​We can’t just generate stubs for all packages in one go (or ever)
  • ​​Many PyPI packages used by few projects at most
  • ​​Some PyPI packages are used by many thousands of users
​​
​​​🤔​ How can we decide which packages are the most useful?

​​Idea: PyPI Download Counts

  • ​​Stats on top downloaded PyPI packages are readily available
  • ​​Problems:
  • ​​Many packages only have high counts because a popular package depends on them
  • ​​What about Conda, deb/rpm packages, and private pip repositories?
  • ​​Big CI jobs
  • ​​VM provisioning

​​Idea: Search Hits on GitHub 

  • ​​Find all from pkg import​ and import pkg​ across GitHub
  • ​​Problems:
  • ​​Identical code copied over many repositories
  • ​​Packages with generic names that are hard to search for
  • ​​Low-value repositories
  • ​​Legacy code and obsolete packages
  • ​​Proprietary code missing

​​Idea: PyPI Downloads + GitHub Search

  • ​​Libraries downloaded N+ times and M+ search hits on GitHub
  • ​​Results look good ​👍​ 

​​Challenge 2: Generating Baseline Stubs

  • ​​Goal: generate baseline stubs for 100 top packages