Ah, the goal makes sense. I would suggest that it’s not so bad to have a controller node fan-out and fan-in queries, as long as the database can handle many concurrent queries. Essentially you’re distributing the work evenly across nodes but you don’t have affinity for a particular node. Yes, there is more latency (it is as slow as the slowest connection) but it is endlessly scalable. But, I am sure I missed some benefits from localizing calculations to only a node or two.

In the scheme above, by the way, it DOES localize searches on one shard. Essentially all relations to a stream are on the same shard as the stream. And each center+radius has one associated stream and therefore the search takes place on one shard.