Does this fix the pandas/scipy issues with NumPy HEAD?

I think the warning is a bit too technical and would be cryptic to many users (including me if I didn't know what the issues was). A gist of the issue would be sufficient in my opinion. Users who are utilizing __array_function__ would probably find their way out themselves. Something along the following would be better since it mentions objects that are familiar to everyone.

              'arrays to stack must be passed as a "sequence" type such '
              'as lists, tuples and so on. Support for non-sequence '
              'iterables such as generators, is deprecated as of NumPy '
              '1.16 and will raise an error in the future.',
              FutureWarning, stacklevel=stacklevel)

@ilayn good point, I adjusted the warning text to your suggestion. I agree that it's probably not necessary to mention __array_function__ here.

@mattip It definitely fixes pandas on NumPy head. I don't have local dev installs of scikit-learn or scipy setup, so I can't confirm for those yet.

SciPy has already merged the compensating changes that remove generator usage in this context so that our CI remains "green."

I restarted the buggy shippable builds here

Why FutureWarning when the message says "deprecated"? Just Deprecate the sucker.

Why FutureWarning when the message says "deprecated"? Just Deprecate the sucker.

FutureWarning is now described as "Base category for warnings about deprecated features when those warnings are intended for end users of applications that are written in Python."

But I guess I'm still used to the old policy: "Changed in version 3.7: Previously DeprecationWarning and FutureWarning were distinguished based on whether a feature was being removed entirely or changing its behaviour. They are now distinguished based on their intended audience and the way they’re handled by the default warnings filters."

I was thinking of this as a behavior change, but I suppose that isn't exactly true -- we will be raising an error instead.

Do we have a NumPy specific policy about when we use DeprecationWarning vs VisibleDeprecationWarning vs FutureWarning?

Do we have a NumPy specific policy about when we use DeprecationWarning vs VisibleDeprecationWarning vs FutureWarning?

Sounds like VisibleDeprecationWarning should derive from FutureWarning at some point in the future. I think this is a consequence of Python hiding DeprecationWarning by default, a change that was a kludge to fix the mistake of issuing a DeprecationWarning in Python 2.7 about some change in Python 3. Or at least, that is how I remember it :) To be clear, VisibleDeprecationWarning is used when we want everyday users to see the warning by default, DeprecationWarning is for downstream developers who may only see it in development testing, and FutureWarning for changed behavior. I suppose we can adopt the Python 3.7 convention at some point, but we will need to have a way to signal changed behavior vs disappearing features.

Sounds like we need to document the use of DeprecationWarning someplace.

Thanks Stephan.