The invention discloses a fault tolerant cluster system and method based on message log. The system can obviate the system overhead produced by recording synchronous log by storing checkpoint and message log using a backup process and simultaneously recording the checkpoint and message log in a memory on a message sender side without increasing extra reliable equipment. This can not only record the log recording cost but also eliminate the dependency on stable storage medium. Without using any reliable storage equipment for storing checkpoint and log, the invention does not depend on extra spare calculation nodes to replace failure node in recovery period, and the process can continuously run on the rest of nodes without rebooting the process. At the same time, the system can also conveniently increase the load balance function to effectively reduce the influence of node failure on the entire system.