-s disables cross-checks between listed files

[toejough]

I'm having a compound issue, and I'm not sure if each part of this deserves an issue. The thing that bothers me the most is the missing cross-checks, but to get there, it might be helpful to explain some context...

I've got a project that:

• uses local, fully typed files
• uses a third-party, fully typed module (call it is_typed)
• uses a third-party, untyped module (call it not_typed)

If I run mypy a.py b.py initially, I get an error about how it can't find is_typed, and maybe I should either use MYPYPATH or --silent-imports.

When I run MYPYPATH="/path/to/virtualenv/site-packages" mypy a.py b.py, I get a ton of failures related to not_typed, plus a few about failures where a.py calls something in b.py with the wrong type.

These last set are the errors I'm calling cross-check failures. I want to see those. I don't want to see the earlier set - I know that not_typed is not typed, and I'm not going to make stubs for that project at the moment.

Back to the example. Setting the full site-packages dir grabbed more than I wanted. If I run MYPYPATH="/path/to/virtualenv/site-packages/is_typed" mypy a.py b.py, I get AssertionError: Neither id, path nor source given, with a stack trace. If I add --pdb to the call to try to debug, I do not get dropped into the debugger.

Ok. Well, let's try the --silent-imports then. mypy -s a.py b.py results in a clean run. The cross-check failures do not show up, even though they're still there. :-(

Ok. Well, let's try from not_typed.sub import thing # type: ignore . Then run MYPYPATH="/path/to/virtualenv/site-packages" mypy a.py b.py again aaaaand... All the errors for not_typed.

So then I got desperate and tried thing = importlib.import_module('not_typed.sub').thing, and that resulted in Invalid Type "b.thing" everywhere I used the thing type in b.py.

I think I ran into several real problems here, but I'm not convinced I'm not just doing it wrong somehow, because I don't see anyone else having the same issues here, and I don't think my use case is all that tricky.

Is there anything I can do differently?

[gnprice]

Thanks for this writeup! I think the shortest direct answer is that we don't currently have great support for third-party libraries that have themselves adopted PEP 484 types -- that is, for libraries that aren't in your own codebase's source tree and live instead in something like a site-packages/, but that have their own inline annotations which you'd like to use rather than finding or writing stubs.

Clearly this is a situation that will become more and more important as more library authors adopt static type annotations -- we should definitely find a better solution than we have.

The workflow that currently definitely does work great is: all the files you want to actually type-check are in your source tree, and you pass them (or the directories containing them) to a mypy command line. All the external libraries you use either have stubs, or they don't have type annotations, and if any are in the latter situation then you pass --silent-imports on your mypy command line.

That workflow is pretty much what you describe with mypy -s a.py b.py, and it should work great -- unless the errors you want to see are dependent on having the actual types from is_typed, in which case you would need stubs for is_typed in order to use that workflow. Is that your situation? If not, there may be a different issue present.

Issue #1190 is the main one we have discussing this problem. Does that look like it covers what you're describing?

[gvanrossum]

The one thing I'd add is that setting MYPYPATH to point to site-packages is basically always wrong, and is likely to cause crashes. This is not great, but it's a good thing to be aware of in the current stage of mypy's development.

[toejough]

@gvanrossum - ah, ok. thanks for the tip!

@gnprice - thanks for the response!

So, maybe I should stop filing bugs at the end of my work day. #1190 is definitely part of the problem I was having, so thanks for pointing that one out.

The bigger part of my problem, though, was that it seemed like -s was hiding errors where a.py called a b.py function with the wrong types. I've made a bunch of changes to my actual code since then (and unfortunately squashed them all down, so I can't get back to where I was when I filed the issue), and now I can't reproduce that condition. I can't reproduce it with a minimal example, either, so I guess I'll close this issue.

My best guess is that the type error I was looking for was actually dependent on is_typed's types, which means that yes, my whole problem is covered by #1190.

Thanks again for the responses and thought - and apologies for the noise. If I can reproduce AND get it down to a minimal example, I'll open a new issue, but for now I'd consider this one closed.

[gvanrossum]

One more tip: -s is a measure of last resort. Also try -almost-silent instead.

…

--Guido (mobile)