Thanks! This is always a bit trickier with regressions in masked arrays and not being quite sure how this is used.

I actually, think a slightly more maximal approach may be nice, but with a release note and no backporting.
If this is important to you to backport, I might try to go more minimal and just skip recursion into strings for now.

The more maximal approach would be to do the following (sorry, it is a bit more complicated, in the sense of changing the pattern!):

1. Rather than flattening each element, flatten the arr.dtype (if it is structured). There is a function that does this in iotools: flatten_dtype.
2. We can just use dtype = np.result_type() on the flattened dtype.
3. We have to create the final array, without moving at least a bit to C, I suspect the easiest thing will be to iterate the fields and assign one by one, like:
   • for i, name in enumerate(dtype.names): out[i+something] = input[name]
   • Unfortunately, subarray fields are a bit tricky and need some thought there.

This approach will also change the result dtype, e.g. for a struct array of float32 where the dtype will now still be float32 (not just for dtypes that include objects as you have it).

The numpy.lib.recfunctions probably has some related functionality, although similar to masked arrays, some of the code may not be something we should adopt exactly.

Thanks! This is always a bit trickier with regressions in masked arrays and not being quite sure how this is used.

I actually, think a slightly more maximal approach may be nice, but with a release note and no backporting. If this is important to you to backport, I might try to go more minimal and just skip recursion into strings for now.

The more maximal approach would be to do the following (sorry, it is a bit more complicated, in the sense of changing the pattern!):

1. Rather than flattening each element, flatten the arr.dtype (if it is structured). There is a function that does this in iotools: flatten_dtype.

2. We can just use dtype = np.result_type() on the flattened dtype.

3. We have to create the final array, without moving at least a bit to C, I suspect the easiest thing will be to iterate the fields and assign one by one, like:

   • for i, name in enumerate(dtype.names): out[i+something] = input[name]
   • Unfortunately, subarray fields are a bit tricky and need some thought there.

This approach will also change the result dtype, e.g. for a struct array of float32 where the dtype will now still be float32 (not just for dtypes that include objects as you have it).

The numpy.lib.recfunctions probably has some related functionality, although similar to masked arrays, some of the code may not be something we should adopt exactly.

Thanks for replying.
Should I implement the maximal approach you mentioned?

@seberg @charris
The last changes skip recursion for strings. Maybe, it can be just used for backporting now.
Then, another PR with the maximal approach mentioned can be done.

#30872 fixes the test failures seen here, a rebase should fix it.

Thanks, let's get this minimal approach in then! FWIW, I think the real fix is still to revamp the approach to walk the dtype fields.