Docs: Static analysis library capabilities

[fkromer]

It understands Python on a deeper level than all other static analysis frameworks for Python. [jedi.readthedocs.org]

Would be great to have more info about static analysis library capabilities to allow comparing jedi with pylint (plugin Interface), astroid, ast, etc.

[davidhalter]

I actually removed that sentence a few days ago. It felt like a strange sentence, for a couple of reasons:

1. It's arrogant.
2. I haven't really compared the solutions in quite some time.

But since we're here already, I can give you some background that we might eventually document somewhere.

There are 4 solutions on the market that offer support for your IDE:

• Jedi
• https://github.com/PyCQA/astroid
• https://github.com/microsoft/PTVS (I think it's this repository)
• PyCharm

The community driven solutions Jedi/Astroid have very different goals for type inference. Astroid tries to analyze entire files and therefore needs to be really fast and therefore sometimes imprecise. This is what Jedi is lacking, it cannot analyze entire files, it's way to slow for that. Not that Jedi is bad software, it just tries really hard to find the correct types. So I would still argue that Jedi understands Python on a deeper level than astroid, but that just comes from the different use case. I haven't really used astroid, though.

PyCharm is clearly the best all-round solution, but they still don't follow decorators and also sometimes make other type inference mistakes. They are really good at working with an index (See #1059). They have in general a pretty solid solution, that PTVS probably tries to imitate. Jedi again is probably still more precise, but Jedi doesn't have all the features: No linting, no refactoring, etc.

For goto/completions, I still think Jedi is probably the best solution, but the other solutions don't just focus on that.

There are 4 static type checkers (using mostly typeshed and annotations):

• https://github.com/python/mypy
• https://github.com/Microsoft/pyright
• https://github.com/google/pytype
• https://github.com/facebook/pyre-check

Knowing the companies behind those solutions, most of these will eventually die and one or two will remain, but in general they serve one purpose: Static type checks. This means that they are working with annotations. If you don't use annotations in your code, these are mostly useless. They are there for specific use cases. So they are not really comparable to the solutions above.

Everything else is either old and unmaintained (e.g. rope), incomplete or proprietary and nobody know about it.

[fkromer]

I did not consider the sentence arrogant. However I evaluated different libraries for python-tia and was wondering about how Jedi is different. My use case is analyzing Python source code (primary Python 3, potentially Python 2 as well) and getting the location of classes, class methods, functions, etc. as file line ranges. It seemed like Jedi is not the best fit for this use case. W.r.t. performance ast seems to be the best fit. But your project parso has a more abstract API.

[davidhalter]

Yeah, parso is unfortunately not as fast as ast. That might change in the future, but will take some time.