A system and method are disclosed for managing a plurality of virtual machines (VMs) in a fault-tolerant and low-latency manner. In accordance with one example, a computer system executes a first VM and a second VM, and creates a first live snapshot of the first VM and a second live snapshot of the second VM. The computer system detects, after the creating of the first live snapshot and the second live snapshot, a failure that affects the executing of the first VM, and in response destroys the first VM and the second VM, creates a third VM from the first live snapshot and a fourth VM from the second live snapshot, and initiates execution of the third VM and the fourth VM. An output transmitted by the first VM to the second VM during the creating of the first live snapshot and the second live snapshot is not blocked.