Curiously, the dot function behaves differently with the oldnumeric version.

In [6]: x.dot(x)
Out[6]: array(10.0)

In [7]: dot(x, x)
Out[7]: 9e+20

In [8]: np.__version__
Out[8]: '1.8.1'

Not really, ma.dot behaves the same in oldnumeric.ma and ma.

>>> numpy.ma.dot(x, x)
masked_array(data = 10.0,
             mask = False,
       fill_value = 1e+20)

Note that MaskedArray.sum properly skips masked values and we have

>>> (x*x).sum()
10.0
>>> x.dot(x)
9e+20

What is the "right" behavior here?

Since ma.dot(x, y) is defined as numpy.dot(x.filled(0), y.filed(0)), there is a strong precedent for having MaskedArray.dot do the same, but ma.dot has this unfortunate behavior:

>>> print x
[1.0 -- 3.0]
>>> print numpy.ma.dot(numpy.eye(3), x)
[1.0 0.0 3.0]

It would be better if masked values behaved as "weak nan" and eye.dot(x) be the same as x. ("Weak nan" has the property nan * 0 = 0.)

I believe this issue was fixed by PR #5709 as indicated in the comments. Can you verify @abalkin?

Also, the same issue seems to have been reported in gh-3219, did the PR fix this also?