I think using a pyarrow dataset as an intermediary would allow for a zero copy read/write from one connection into a second connection.