API: .to_numpy()

[jreback]

xref pandas-dev/pandas#14052

currently we have an (implicity) numpy conversion when we access .values of a 1D (Series). This mostly returns a numpy array, though we do return numpy-like objects several dtypes:

• categorical, we simply return a Categorical object
• datetime tz/aware, we return a datetime64[ns] array in UTC (losing the tz)

This also has implications when we have a 2D object (DataFrame). we use a type that can safely hold all of the data:

• int & floats -> floats
• datetime w/tz -> object array
• object & anything -> object array

so generally this is ok for 2D in that you preserve as much as possible (though of course you must copy / return heavyweight object array at times).

So need some though on how to make this api look & validate cases.

I would propose .to_numpy() (a function, so we can potentially pass options). and it won't break the current API (which we can preserve I think / provide back-compat). w/o making libpandas jump thru hoops to support the 'old' stuff.

[wesm]

I agree with this -- it would be helpful to start migrating away from the .values API toward something more explicit to ease the burden. We might even want to introduce a logging layer into pandas 1.0 to alert users to use of "non-future proof" APIs

[jreback]

right I suppose could instrument things maybe via an option

pandas.options.future.logging='warn'|'raise'|'ignore'

and namespaced a bit for .future.* options if we need.