Sorry, I need to correct that. I meant that shards are basically still primary / secondary with an arbiter for failover. We prefer horizontal nodes instead, like Cloudant, Elastic Search or Cassandra offers.
Space/resources was certainly a concern in this case. A disk-based solution with predictable latencies enables storage of far more data than something that requires holding the working set in RAM.
I don't quite understand this, that architecture was basically deprecated a while back, can you elaborate?