That sounds odd, since if it was sharded by term, then a given search would only need to go to a handful of servers (one for each term) and then the intermediat

That sounds odd, since if it was sharded by term, then a given search would only need to go to a handful of servers (one for each term) and then the intermediate result combined. But with it sharded by document, every query has to go to all the nodes in each replica/cluster. Exactly. Then a flash crowd occurs and your shard fails to service requests.