I am not sure if this true for current hardware, but once upon a time the mainframe CPUs contained duplicates of each functional unit that ran in lock step; after every computational instruction, the results of the units are compared; if they differ, the CPU repeats the instruction once. If the results
still differ, it "calls in sick" to the operating system.
The OS then brings a spare CPU online and transfers the program that was running on the failing CPU to the new CPU, takes the failing one offline, and, depending on your service contract with IBM, calls home for a replacement CPU. The program does not even notice something went wrong. The next day an IBM service technician rings your data center's doorbell and replaces the faulty CPU, all without taking the machine offline.
That kind of resiliency and redundancy runs throughout every aspect of the system's design.
If you can afford it, having a mainframe be your single point of failure is not too bad.
The OS then brings a spare CPU online and transfers the program that was running on the failing CPU to the new CPU, takes the failing one offline, and, depending on your service contract with IBM, calls home for a replacement CPU. The program does not even notice something went wrong. The next day an IBM service technician rings your data center's doorbell and replaces the faulty CPU, all without taking the machine offline.
That kind of resiliency and redundancy runs throughout every aspect of the system's design. If you can afford it, having a mainframe be your single point of failure is not too bad.