Therefore now applied on eg the integer columns separately. In a test case, this caused the integers to overflow for prod(). While before, the DataFrame was converted to a float ndarray (because it also has a float columns), not giving those overflow issues. Hence the test was failing (see test dask/dataframe/tests/test_ufunc.py::test_ufunc_with_reduction). The question here is whether dask wants to follow that behaviour.