This makes it critical that requests get dispatched to nodes that will have the data cached (if such a node exists), especially on Lookup requests which have lots of overlapping subproblems.