A fault tolerance method and system for VMs on a cluster identifies a client state for each client session for those applications. The method replicates the client session onto a primary and a backup VM, and uses a network controller and orchestrator to direct network traffic to the primary VM and to periodically replicate the state onto the backup VM. In case of a VM failure, the method reroutes network traffic of states for which the failed VM serves as a primary to the corresponding backup, and replicates states without a backup after the failure onto another VM to create new backups. The method may be used as part of a method or system implementing the split/merge paradigm.