Smart Client

For single-shard queries:
- A client issues the query on the coordinator.
- The coordinator finds the right node and sends the query to that node.
- The query gets executed on the node.
- The result is fetched to the coordinator.
- The result is returned to the client from the coordinator.

If clients knew where the query should be executed the flow would be:
- The client issues the query on the node that has the shard.
- The query gets executed on the node, with local execution, avoiding an extra connection.
- The result is returned to the client from the node.

So, we would avoid the extra round trip from coordinator to the node and would also distribute the load a bit.

We can have a client side library, to make clients “smart clients”, which know what node has the data for a given value.

We should do some performance test to see what would be the potential performance benefit of `smart clients`.