Yes... I think there is an older bug report about this... The basic problem is that somewhere the masks shape is assigned using mask.shape = new_shape (along those lines, in array_finalize or so), which fails to create a new view, and thus can change the mask of the earlier one as well. A long time ago I had a short look at this, not sure if it is quite right, but maybe you can have a look, do necessary cleanups and make a pull request. Here are the changes from back then:

https://github.com/seberg/numpy/compare/safemask

I did change the .shape attribute completely, I am not 100% sure as to why all those changes though.

So the more thorough fix would probably include changing the ndarray.shape attribute setting logic (possible fixing #145 while at it. I am not certain as to the discussion, but since I messed with it recently, attempt_nocopy_reshape can easily be modified to cover all cases (if someone works on that, I have had done it, but it would be unused right now so it is not in numpy), so if it is exposed to the getset.c it could be just called directly then.

But just saying if someone feels like looking at into finding a good solution. I don't think its difficult overall, and I am happy to give pointers.

@seberg, thanks for a quick and thorough reply.

I'm a bit confused right now. Do you think working on https://github.com/seberg/numpy/compare/safemask would be a good way to fix this or do you think this issue should be tackled by fixing #145? The latter is harder to understand for me and I don't know why it hasn't been merged yet?

Also, I haven't contributed to Numpy yet, so I'm not very familiar with the workflow (though I've read the development workflow). Is this easy enough for a first contribution?

Don't know, but I had a bit longer look, and I think fixing the latter (gh-145) has some tedious details, and maybe the approach in that PR is best. I just though that by changing ndarray.shape attribute setting, you may not need to give the masked array its own shape attribute. While I do not remember why I did that, there was probably some reason...

So my guess is, the basic approach in that safemask branch is probably good, but needs at the very least some tests.

Contributing is very easy, all you need to do is create your own branch in a github fork, push to it and then github will already show you a button to create a pull request. And once you have the pull request open, anyone can easily comment on it to smooth out the details.

The bug is still present identically in recent numpy versions.

>>> np.__version__
'1.14.3'
>>> A.mask.shape
(1, 2)
>>> A.shape
(2,)
>>> A
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/Users/honnorat/usr/conda/lib/python3.6/site-packages/numpy/ma/core.py", line 3942, in __repr__
    self._insert_masked_print(),
  File "/Users/honnorat/usr/conda/lib/python3.6/site-packages/numpy/ma/core.py", line 3866, in _insert_masked_print
    _recursive_printoption(res, mask, masked_print_option)
  File "/Users/honnorat/usr/conda/lib/python3.6/site-packages/numpy/ma/core.py", line 2439, in _recursive_printoption
    np.copyto(result, printopt, where=mask)
ValueError: could not broadcast where mask from shape (1,2) into shape (2)

Thanks for reviving this - I think you just found the root cause of #10270. Pinging @mhvk, who has a PR somewhere that is a smaller alternative to #10308, that was just waiting on a test.

Here's that test:

np.tile(np.ma.masked, (2, 1))
assert_equal(np.ma.masked.mask.ndim, 0)  # fails

@honnorat - thanks, indeed. My PR 1308 #10314 indeed solves this problem, and now uses @tjanez's example on top in a test that should ensure the problem will not recur.

@mhvk: Wrong PR number?

Thanks, corrected to #10314

Hello, I have run the snippet of code in the description and I am under the impression that this issue can be closed?

Python 3.10.7 (main, Sep  7 2022, 00:00:00) [GCC 12.2.1 20220819 (Red Hat 12.2.1-1)] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import numpy as np
>>> np.__version__
'1.24.0.dev0+934.gdb7414b7f'
>>> A = ma.masked_array(np.array([1., 2.]), mask=np.array([False, False]))
>>> A
masked_array(data=[1.0, 2.0],
             mask=[False, False],
       fill_value=1e+20)
>>> B = np.tile(A, (2, 1))
>>> A
masked_array(data=[1.0, 2.0],
             mask=[False, False],
       fill_value=1e+20)
>>> A._data
array([1., 2.])
>>> A._mask
array([False, False])

Thanks for listening.

Thanks for the note @cmarmo. Agree, seems this has been fixed years ago! (in gh-10314)