Doing the same operation using plain numpy, the numpy_nullable backend and the pyarrow backend shows that numpy_nullable behaves differently than the other two and in a rather counter-intuitive way. >>> import pandas as pd >>> sr = pd.Series([1.5, 0.0]) >>> (sr / sr).max() 1.0 >>> sr = pd.Series([1.5, 0.0]).convert_dtypes() >>> (sr / sr).max() <NA> >>> sr = pd.Series([1.5, 0.0]).convert_dtypes(dtype_backend='pyarrow') >>> (sr / sr).max() 1.0 Is there a good reason, or at least an explanation, for this inconsistent behavior? Is there a way around it or should we simply avoid the numpy_nullable backend? Continue reading...
