As part of the new rs->configShard procedure, there is a period of time where users are connected directly to a replica set started with --configsvr, but the replica set has not been added as a shard yet. As a result of this, TTL index delete will fail with NamespaceNotFound because it will attempt to get sharding database metadata which is only created after the replica set has been added as a shard. 

A possible solution could be to attempt to refresh the sharding filtering metadata on namespace not found error similar to how StaleShardVersion errors are handled.