Our current monolithic approach to Python packaging isn't likely to be sustainable long-term.
At a high level, I would propose a structure like this:
We can maintain the semantic appearance of a single pyarrow package by having thin API modules that would look like
Obviously, this is more difficult to build and package:
- CMake and setup.py files must be refactored a bit so that we can reuse code between the parent and child packages
- Separate conda and wheel packages must be produced. With conda this seems more straightforward but since the child wheels depend on the parent core wheel, the build process seems more complicated
In any case, I don't think these challenges are insurmountable. This will have several benefits:
- Smaller installation footprint for simple use cases (though note we are STILL duplicating shared libraries in the wheels, which is quite bad)
- Less developer anxiety about expanding the scope of what Python code is shipped from apache/arrow. If in 5 years we are shipping 5 different Python wheels with each Apache Arrow release, that sounds completely fine to me.