@wesm Is there any reason to use _register_pandas_arrow_handlers over _register_custom_pandas_handlers? Why not always use _register_custom_pandas_handlers? Is there a difference in generality?

I'd be perfectly fine with the faster pandas serialization code path. I guess the one thing separating the pandas_serialization_context from the default one would be the handling of NumPy object arrays?

Is that related to the choice of _register_pandas_arrow_handlers versus _register_custom_pandas_handlers? Both of those can be used with both of the custom numpy object array serializers, right?

EDIT: I think I see what you mean, yes, if we get rid of _register_pandas_arrow_handlers, then the only difference would be the handling of numpy object arrays.

This seems like the obvious thing to do unless there is some drawback with _register_custom_pandas_handlers.

Yes, thanks @devin-petersohn! This PR looks good to me.

@wesm, is speed the only difference between the two implementations? Just curious if there are edge cases.

I think the new version will have marginally better support for "weird pandas data". If you run into issues please open issues so we can investigate