When the standalone config is empty, e.g., after an error or after stopping the server, on the next insert operation, Cluster#next_primary returns nil and causes errors up the call chain. The cluster.service.scan! call path demonstrates that each server config is updated by the results of the ismaster command, or empty {} for any error.

The following is a possible fix.

  class Cluster
    def next_primary
      if client.cluster.mode == Mongo::Cluster::Mode::ReplicaSet
        primary = client.server_preference.primary(servers).first
      else
        primary = servers.first
      end
      raise Mongo::NoMaster.new("no master") unless primary
      primary
    end
  end

There's probably a better way to fix this via the server_preference mechanism, but at least the above should help to clarify the problem.