What's the design decision regarding whether operations can return another backend's object?

[anmyachev]

Let's assume we have a pyarrow-backed dataframe/series. Can any operation return numpy-backed dataframe/series, or should this never happen (and if it does, it will be treated as a bug)?

I didn't find an answer to this question here:
https://pandas.pydata.org/pandas-docs/version/2.2.2/user_guide/pyarrow.html

[mroeschke]

Thanks for the question. I would say generally an operation with a specific a dtype backend (pyarrow, nullable numpy, normal numpy) should return the same dtype backend unless

1. An API states it should return a specific type, e.g. ExtensionArray method APIs
2. Mixing two different backends in an operation, e.g. int64[pyarrow] + Int64. (Probably not well defined in all cases)

[WillAyd]

This is a great question and something that it would serve us to best define in the future though. For simple types like ints, floats, etc... where NumPy and pyarrow share the same storage layout (at least for the data buffers), I think NumPy + Arrow should return Arrow; otherwise it would be a lossy operation

For more complex types that's an open question, but I would still prefer for Arrow types to come out on top