We have collocated Spark and Cassandra nodes and observe that just a subset of created tasks are running on locality level node_local when reading a Cassandra Table with SCC and performing a count on it. All shuffle read related tasks are running node_local, but shuffle writes are running on locality level any. Is this behaviour as expected? When replacing count with collect, we observe only tasks with locality level any. Shouldn't transformations that do not trigger a shuffle also create tasks with locality level node_local?