yeah in the pythonnet context Arrow still makes more sense than pushing everything through PyBuffer.
PyBuffer_GetBuffer works and it’s totally fine for simple contiguous numpy arrays. But once you move to tables, pandas data, nullable columns, strings, etc, you end up manually handling layout, dtype, nulls and lifetimes on the C# side. pythonnet doesn’t really remove that complexity.

arrow is basically built for this exact problem:

• python produces data
• c# consumes it
• clear schema instead of raw pointer + length
• zero-copy or minimal-copy options

typical setup with pythonnet:

• python side: use pyarrow to create Table / RecordBatch
• transfer via Arrow IPC (even in-memory stream works well)
• c# side: read using Apache.Arrow

this integrates cleanly with pythonnet and avoids abusing the buffer protocol.

If you really need true zero-copy, Arrow C Data Interface can be used from pythonnet too, but lifetime management becomes tricky (Python GC, ownership rules, etc). For most cases IPC is simpler and safe.

so overall:

• raw numeric buffers → PyBuffer is fine
• structured/tabular data → Arrow is the right abstraction
  arrow isn’t extra overhead here, it’s the thing that actually makes Python ↔ C# data sharing sane.