In order to fault-tolerantly perform an application process in a server cluster, the process is replicated on multiple servers. In this paper, we consider the replications for the online learning. Here, not only the process but also the replicas of the process obtain the models for the machine learning. The more replicas of the process are performed, the more servers work to perform the replicas. This means, the more electric energy is consumed by servers in a cluster. In order to reduce the electric energy, computation of replicas has to be reduced. Hence, we newly propose the replication schemes for the online learning in this paper. We assume that a score for a model is obtained based on the metrics of the model. Suppose clients can use a model if the score of the model is bigger than a threshold. Some function of a primary which is not performed in a secondary is decided based on the threshold. A primary learns all the data which arrive at the primary. On the other hand, a secondary stops to learn the data if a score of a model obtained in the secondary is bigger than the threshold. Therefore, computation in a secondary is reduced.