You won't need to create a client at all in the node. Ideally nodes should only contain any transformation logic. The dataset abstraction is available so that you can just change the destination/source filesystems without having to change your pipelines at all.