Happy to do this. We could also go through an additional VisibleDeprecationWarning cycle, but I suppose, we always have the option to revert the removal and do that instead...

Lets sign it off in the meeting today (technically, its already accepted so...).

Another option: replace the functions with stubs that just raise RuntimeError, just so that users get a better message than AttributeError

Hmm, the error idea is indeed maybe a bit gentler and only leaves like 50 lines of code probably. Too bad we can't put it directly on numpy.financial yet.

More than one person suggested using the module __getattr__ function to handle the error message. I pushed a commit that implements this by adding __expired_attributes__ in numpy/__init__.py. It is handled similar to __deprecated_attributes__, but raises an AttributeError with an informative message instead of a warning.

For example,

In [1]: import numpy as np                                                                           

In [2]: np.npv                                                                                       
---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
<ipython-input-2-ab016edb79d9> in <module>
----> 1 np.npv

~/mc38/lib/python3.8/site-packages/numpy-1.20.0.dev0+2558224-py3.8-macosx-10.9-x86_64.egg/numpy/__init__.py in __getattr__(attr)
    245             try:
    246                 msg = __expired_attrs__[attr]
--> 247                 raise AttributeError(msg)
    248             except KeyError:
    249                 pass

AttributeError: In accordance with NEP 32, the function npv was removed from NumPy version 1.20.  A replacement for this function is available in the numpy_financial library: https://pypi.org/project/numpy-financial

In Python 3.6, a plain AttributeError will be raised, but we will likely drop (official) support for 3.6 in the 1.20 release, so maybe that's OK.

This adds a bit more overhead when accessing the numpy namespace, so perhaps this isn't something we want to do.

@WarrenWeckesser thanks, it doesn't actually add additional overhead. The __getattr__ is only called if the item is not already found in the module dictionary. Or yes, it adds overhead, but only when attribute lookup fails :).

@seberg, thanks, that's good to know.

Thanks, Warren, pretty amazing to see this actually finished! Now I just wonder how much fallout we will see ;).