@abalkin Test failures are universal. You should be able to check at home and fix them up.

Travis is green now. a49c279 illustrates possible backward-compatibility problems.

The documentation of the function needs fixing/updating. It isn't clear to me that the function should always return a new array, doing so may break current code. What is the argument for this change? That argument should also be in the commit message.

@charris - please see discussion at gh-4039. In short, ma.MAskedArray constructor is very slow and traditionally asarray functions would pass through objects of the correct type. This is what both numpy.asarray and numpy.oldnumeric.ma.asarray do.

As such, I think it would be correct to just pass through a masked array. But maybe we should ping the mailing list about it. It could potentially create some backward compatibility problems, afterall. Needs a prominent warning in the release notes in any case.

Looks like it should already not be making a copy unless needed

return masked_array(a, dtype=dtype, copy=False, keep_mask=True, subok=False)

If it is slow, or not working correctly, maybe masked_array needs some work.

LGTM, but I notice that ma.asarray has an order keyword which is not used, nor available in masked_array. Also, ma.asanyarray doesn't have an order keyword, but is documented as having one. Could you fix that up before this goes in?

Or rather, I'll put this in to encourage you ;) And then could you fix the documentation/functions?