The two major approaches to building fault-tolerant services are commonly know as the Primary Backup approach (PB) and the State Machine approach (SM). PB can tolerate crash and omission faults and runs more economically than SM, but SM can tolerate more serious faults, including arbitrary or Byzantine faults. Instead of selecting one or the other approach, this either incurring a high running cost or risking the service becoming incorrect when unexpected faults occur, we advocate the approach of adaptive fault tolerance. We present algorithms that intelligently adapt between PB and SM, thus retaining (almost) the best of both worlds. Our adaptive approach is modular in that any PB or SM protocol can be used, and is also practical in that it can be easily incorporated into some existing systems.

• sri-csl-tr94-04.pdf
• sri-csl-tr94-04.ps