Yeah, I‘ve also seen the case where you start with Spark but at some point need to move to Pythonland and hold your data in-memory and it doesn‘t scale. TF is much more naturally able to deal with large amounts of data.