Also, there is a test that is creating a ragged array and emitting a warning, but the warning somehow slips under the radar of pytest, maybe because it is inside an assert_raises

The solution to the first problem is to pro-actively create the array with an object dtype:

>>> import numpy as np
>>> arr = np.array(["a", "b"], dtype=object)
>>> objs = np.array([arr, "foo", arr], dtype=object)  # HERE
>>> np.sum(objs, axis=0)

I am probably reacting to an overly-simplified example since this is a snippet from a more complex code base.

@jbrockmendel is using np.sum(np.array(objs, dtype=object), axis=0) in some way undesierable for you? (or even going further and using arr = np.empty(len(objs), dtype=object); arr[:] = objs to not rely on numpy guessing the right thing at all)

@mattip @seberg thanks for the quick feedback. In the np.sum with strs case, the suggested workaround likely works. To get an idea for where else this comes up in the pandas codebase, see pandas-dev/pandas#30035

The change here https://github.com/pandas-dev/pandas/pull/30035/files?file-filters%5B%5D=.py#diff-d5ffc8f19bc8362967ce9ddd23026739R141 was needed to silence warnings caused by calling operator.mod (i dont remember what the inputs were off the top of my head)

Closing the correct fix is to explicitly create the ndarray before calling a ufunc.