Can't do this until state moves from per connection to per reqeust.
when lots of web server, there is too many connections
Mongos crashes by few seconds of DB locking (or network delay).
Remove WriteBackListener entirely, replace with batched writes to shards
Make mongos connection pool more configurable for users with large #s of shards
Refactor how shard version information is associated with each connection
Enable mongos mode which releases ShardConnections to the pool after read ops
Document multiplex improvement in 2.6 release notes
Cap number of connections between mongos and mongod