If it's always just numpy.asarray, IMHO we should put that at the beginning of the relevant functions. Doesn't cost anything if it's already an array. Telling people to always manually convert is a bit pedantic. We should change our policy to

All of plotting functions expect `numpy.array` or `numpy.ma.masked_array` input, or objects that can be converted to these using `numpy.asarray` (such as `pandas.DataFrame` and `numpy.matrix`).

Well, we can't launder everything though asarray because units. Lack of support for e.g. np.matrix was explicitly added in #3394.
Even if we were to change this policy, this part of the docs would need to stay until we have asarray() calls (+ whatever unit handling needs) in place everywhere.

I see. Then I'd leave this as is. IMHO df.values is more canonical than np.asarray(df). It's shorter to type (not much but still less annoying when working interactively), and people will not necessarily have numpy imported when working with pandas.

Actually to_numpy is preferred over values since pandas 0.24. We should advertise that.

But we support older versions of pandas and I don't think we should start having to field questions from confused users using old pandas. (I don't think the usage tutorial should become a pandas tutorial either.)
I personally like asarray() because it always works (and that's what we use internally anyways); I guess we could just remove the examples here. Or just close the whole PR if you prefer, I'm not overly attached to it.

I'd leave as is; primarily telling "convert to numpy", plus "here is an example" is the right priority.

Anyway, thanks for thinking about how to improve the docs!