We solved this by manually updating the workers to the latest, and then updating the leader, rebooting, and voila.