Looks to be a duplicate of #3126.

This is a bug in ndarray, nor recarray. For example:

>>> a = np.zeros(1, dtype=[('x', float, 5)])
>>> a[0]
([0.0, 0.0, 0.0, 0.0, 0.0],)
>>> a[0]['x']
array([ 0.,  0.,  0.,  0.,  0.])
>>> a[0]['x'] = 1
>>> a
array([([1.0, 0.0, 0.0, 0.0, 0.0],)], 
      dtype=[('x', '<f8', (5,))])

In particular, I think it is a bug in ndarray.setfield bwing too permissive. Maybe this is what the user should be writing:?

>>> a[0]['x'][:] = 1

I think ndarray.getfield and ndarray.setfield are buggy (eg, see the segfauly I cause in #3256 (comment)). I think it might be possible to make them simply call np.view in the right way to solve all these problems. But it may need to wait until some other issues with views are solved (see #5508 #3256 #3286 #2599)

@ahaldane, I have proposed a possible solution to this (raise at least an error) in the pull request #4556 --- which is still under review...

Does this really have to wait? It's just a 15 line long change / patch in production code.

Oh, I missed that link to your pull request above.

I took at your pull request and it seems like a good fix for now. I tested it and it worked, and the code looks OK (though I am a numpy code novice).

What I mentioned about waiting is that I think this problem (plus others) might be solved by making getfield and setfield go through the 'view' machinery. This way all the checks that happen for views would automatically happen for get/setfield.

In #4556 you gave this example:

>>> x = np.empty(1, dtype=[('field', 'i4', 10), ('b', 'O')])
>>> x[0]['field'] = 4 * ones # doesn't work properly currently

I imagine setfield could do something like:

>>> t,o = x[0].dtype.fields['field']
>>> xv = x[0].view(dtype({'names': ['field'], 
                          'formats': [t], 
                          'offsets': [o], 
                          'itemsize': x.dtype.itemsize}))
>>> #then do array_assign(xv, 4*ones)

Taking a view would also solve the following problem:

>>> x[0].setfield(1, 'i8', x.dtype.fields['b'][1])
>>> x
zsh: segmentation fault (core dumped)

Now I admit I have not looked deeply into the assignment machinery yet so maybe this isn't possible...

In my opinion your pull should be merged for now. I mentioned "waiting" only with repsect to trying out these additional ideas about views.