The holy grail of querying.

• Indexes are never shared.
• Each node maintains local indexes for state it is responsible
  for (-Dinfinispan.query.indexLocalOnly=true).
• Indexes could be in memory or disk.
• Queries themselves are distributed.
• The query object is built and broadcast to the entire cluster.
• Each node executes the query on its own local index, returning
  results.
• The calling node returns a CacheQuery impl that lazily fetches
  and collates results from the cluster.
• I expect this Map/Reduce model to perform very well since the
  workload is split up and happens in parallel across multiple CPUs
  against much smaller (individual) datasets.
• Works with all cache modes, including DIST.
• Need to make sure duplicates are handled, as well as failover.