The current implementation uses a 4MB buffer to perform the read. It would be nice if this were configurable via an option as is the case in the rest of the API (you can control chunk size, batch size, and read/write buffer sizes on other methods).