Since in the relatively near future, one will be able to do non-trivial analytical operations and query processing natively on Arrow data structures through pyarrow, it does not make sense to require users to always install NumPy when they install pyarrow. I propose to split the NumPy-depending parts of libarrow_python into a libarrow_numpy (which also must be bundled) and moving this part of the codebase into a separate Cython module.

This refactoring should be relatively painless though there may be a number of packaging details to chase up since this would introduce a new shared library to be installed in various packaging targets.

Reporter: Wes McKinney / @wesm

Related issues:

• C++ project doesn't need to require Numpy with ARROW_PYTHON=ON because it only builds Python extension's support libraries (is related to)

_{Note: This issue was originally created as ARROW-8998. Please see the migration documentation for further details.}