Personally I'd prefer pure functions to methods on the types. I believe this will result in less code to maintain and make argument polymorphism (function arguments that can be different data types) more obvious/discoverable. I don't mind having a few conveniences like `arr.cast(...)` but there's definitely some code bloat from maintaining these implementations for both Array and ChunkedArray
On Tue, Dec 17, 2019 at 3:27 AM Maarten Breddels <maartenbredd...@gmail.com> wrote: > > Hi Uwe, > > Having it in a separate package/module/namespace makes it easier to make it > an optional install in the future, might that happen. Also, it would be more > tab completion friendly. > > Cheers, > > Maarten > > > On 17 Dec 2019, at 10:24, Uwe L. Korn <m...@uwekorn.com> wrote: > > > > Hello all, > > > > we have developed quite some compute kernels in C++ nowadays and I would > > like to call them from Python. We could expose the kernels on the > > Array/ChunkedArray classes themselves or as standalone functions (or as > > both). What would be the preferred way? > > > > Also exposing them as standalone functions bears the question whether to > > have them in the top-level namespace (e.a. pyarrow.sum) or in a separate > > one (e.g. pyarrow.compute.sum). > > > > Uwe >
