The idea is to return the masked singleton, a scalar MaskedArray. The issue here is that masked has currently a dtype(float64). As mentioned in another conversation, we could try to use dtype(bool_) instead (which might however) lead to some problems of precedence and check whether we're not breaking anything w/ matplotlib before considering this PR (the change is a bit too radical to my taste).

Oops, I missed the fact that is used as a singleton... Sounds like this is complicated and likely not much can be done.

@seberg - the traditional use for masked singleton is to check for masked items:

>>> x = np.ma.masked_all(3, dtype='h')
>>> x[1] is np.ma.masked
True

This use case is unaffected by my change.

I find it really hard to justify the following behaviors:

>>> x = np.ma.masked_all((), dtype=bool)
>>> x.dtype
dtype('bool')
>>> (x & True).dtype
dtype('float64')
>>> x.max().dtype
dtype('float64')
>>> (~x).dtype
dtype('float64')

I will run matplotlib tests with the change. I seriously doubt that anyone is using if x is masked test for the result of a ufunc instead of if x.mask or if x.mask.all().

Here is another wart:

>>> bool(x.sum() is np.ma.masked)
True
>>> bool(x.sum() == np.ma.masked)
False

Sorry, I really don't know this well. As far as I can see, this adds one wart to fix another and since I don't really use masked arrays, I certainly don't know if that may be worth it.

@pierregm Can you comment as to what you think we should do with this?

I'm going to close this. A proper solution would probably be to have masked scalars of all types. That option could use some study, especially preserving backward compatibility.

Or possibly we could fudge things, adding a type attribute to the current masked scalar or some such nonsense.