How do you minimize/optimize data movement?

I'm not sure best how to phrase this question, so let me start with a concrete example:

Let's say I have a timely dataflow computation spread over 2 machines. I have records uniquely identified by (foo, bar, baz) tuples. In the dataflow graph, I first use exchange to group by foo and bar together, and then later I need to group by just bar. Ideally I'd like to minimize data movement across machines. So if I'm grouping by foo and bar, ideally every unique bar ends up on one of the 2 machines so that grouping by bar later requires no movement across machines.

Does that make sense? Is that type of optimization supported?

Start by reviewing the exchange operation and the dataflow grouping behavior described in the example. Determine whether grouping by (foo, bar) can preserve placement for a later grouping by bar across two machines. Done means establishing whether this optimization is supported and, if not, defining the required behavior clearly.