The idea of organizing replicas differently to effect secondary indexing is old, it is used in some commercial data warehouse products as I recall. The reason you don't see it used very often in practice is that it tends to scale-out poorly, particularly if the workload is mixed.

Spatial decomposition to effect secondary indexing (HyperDex kinda sorta does that) performs relatively poorly at smaller scales, as would be expected in theory, but in very large scale-out systems the selectivity increases and converges at the limit on what you would expect if every dimension was indexed, minus the high overhead and poor performance of secondary indexing at that scale.

The performance crossover point between secondary indexing and spatial decomposition is somewhere within a couple orders of magnitude of a terabyte of data in my experience, depending on the data and workload. It is difficult to measure these effects across unrelated implementations because orthogonal storage engine characteristics tend to dominate measured performance. If you measure it in-memory on large clusters in a pure way (remove the unrelated database-y stuff), the scaling and efficiency curves are much more obvious.