Comment by rudhdb773b

If the focus is performance, why use a separate process and have to deal with data serialization overhead?

Why not a typical shared library that can be loaded in python, R, Julia, etc., and run on large data sets without even a memory copy?