Partition key is really only of use in a data set when it is being used in a Data Flow that is processing batch reads.  The Partition Key allows data flows batch processing to be distributed vertically and horizontally within is multi-threaded / multi-node system.  In such environments, you will typically have several threads (typically about 5) processing work on any given node to allow for vertical scaling of the work.  If the batch job is to run over multiple nodes for horizontal scaling (pretty typical for large batch jobs), then you will have that many nodes trying to help process the work.  Lets say you have 5 nodes, this would mean there is the possibility of 5-nodes x 5-threads to concurrently work on processing your job.  Each of these should be given a slice of the work and the right way to slice it up is to give the system a key on which to do so.  This is where the partition key plays a role.  Best practice starting points for distinct partition numbers is about 100, but each systems volumes and deployment topology and use case might warrant examination of what that partition key number should be.