ENH: Add warning in median, percentile and quantile when array is MaskedArray

[mproszewska]

Closes #14716

[eric-wieser]

I'm not a huge fan of this approach - numpy.ma is supposed to be built on top of numpy, not integrated into it. Perhaps practicality beats purity here though.

[mattip]

Another approach would be for MaskedArray to use __array_function__ and override the calls with its own implementation

[eric-wieser]

Even if that implementation does nothing but warn that the mask is not supported.

Problem is, __array_function__ is as I remember it all or nothing - we can't add it for one function without breaking everything else.

[mattip]

Problem is, array_function is as I remember it all or nothing

It dispatches on all functions, but we can defer on ones we do not wish to override. The example below this link shows that design pattern

[rgommers]

Problem is, array_function is as I remember it all or nothing

It dispatches on all functions, but we can defer on ones we do not wish to override. The example below this link shows that design pattern

It doesn't, that returns NotImplemented which in turn will raise a TypeError in the __array_function__ machinery. It pretty much is all or nothing as @eric-wieser says. You could explicitly coerce the masked arrays to ndarrays, and then call the np.somefunc again, but that's a pretty invasive change for the purpose of adding a warning.

[charris]

There was some discussion of making a new masked array implementation that would use the __array_function__ option. I think there was even a rough prototype.

[mattip]

Here is the prototype https://github.com/ahaldane/ndarray_ducktypes/blob/master/MaskedArray.py

[eric-wieser]

This is one of those cases where the NotImplementedButCoercible proposal that @hameerabbasi was proposing would be useful. Another middle ground might be to have ndarray.__array_function__ allow subclasses, not just ndarray.

[mproszewska]

Probably the best option is to make a new masked array implementation as in prototype and add warnings/implementation there.
I agree that current approach in this PR is not the best, but I still think that there should be an information about unexpected behavior of median, percentile and quantile functions called on MaskedArray. Maybe just in documentation.
Should I remove warnings in this PR and add appropriate comments in docs?
Could I work on a new masked array implementation? If so, how can I do that?