Consider the following:

>>> import numpy as np
>>> a = np.random.random((3, 2))
>>> b = np.ma.masked_array(a, mask=[[False, True], [True, True], [False, False]])
>>> b
masked_array(data =
 [[0.110804969841 --]
 [-- --]
 [0.955128477746 0.440430735546]],
             mask =
 [[False  True]
 [ True  True]
 [False False]],
       fill_value = 1e+20)
>>> b.mean(axis=0)
masked_array(data = [0.532966723794 0.440430735546],
             mask = [False False],
       fill_value = 1e+20)

>>> b.mean(axis=1)
masked_array(data = [0.110804969841 -- 0.697779606646],
             mask = [False  True False],
       fill_value = 1e+20)

Note that for masked arrays, they also do not return 0.0 for completely masked rows:

>>> b[1].sum()
masked

I am not saying that this is the way it should be, but if there is going to be a difference, then this should be made very clear for former np.ma users.

I've added a section 'Differences with numpy.ma' which uses your example to show how things behave differently. Thanks!

I've added another commit which makes the 'align=' struct dtype flag stickiness better, and fixes ticket #1912. Are there any more changes you think are necessary before this gets merged?

I haven't had time to finish looking through it, but I suspect I won't have anything substantive to add. However, I've been thinking the name PyArray_SetBase is too generic. What is the Base and what does setting it do? So maybe another name and some explanation of the intended use of the function would help.

Sure, the name PyArray_SetBase is what it is because it matches with PyArray_BASE.

I think the possible confusion is due to poor naming conventions in the NumPy API. First, Py* should be reserved for CPython, not NumPy. I think a transition to Npy* should be done when the API has matured some more. Second, PyArray_* has all the PyArrayObject-specific functions and random other functions mixed in together.

The overloading of the BASE property together with the UPDATEIFCOPY flag also seems unfortunate, as it significantly changes the meaning of that property depending on the flag.

I'm guessing that PyArray_SetBase sets a pointer to the array from which a view is taken, so maybe something like PyArray_SetBaseArray or PyArray_SetRootArray, or some such. I really think all arrays should be views of some separate memory/buffer object, but that is another issue.

Can you expand on the interplay between the BASE and UPDATECOPY flags? I don't think it matters for this commit, but I would like to learn something ;)

Yes, except the base can be any object that owns the memory, not just an array. PyArray_SetBaseObject? I agree with you about the separate memory/buffer object.

When the UPDATEIFCOPY flag of arr is set, arr's base is then the original array from which a copy is made. Arr's base is set readonly as well. When arr is freed, it copies its data into its base and sets arr to writeable again. Quite different from arr's base owning the memory...

PyArray_SetBaseObject sounds good, or maybe PyArray_SetBufferObject or PyArray_SetMemoryObject.

The UPDATEIFCOPY sounds like two paradigms getting mixed together.

I'll switch it to PyArray_SetBaseObject. Changing it to a name other than Base would need a bigger cascading set of changes.

I've changed the name to PyArray_SetBaseObject, and added some documentation about it.

Chuck, did you have anything else you'd like changed before merging this?

No, go for it. Might clean up the trailing whitespace though ;)