Specifying graceful degradation in distributed systems

Distributed programs must often display graceful degradation, reacting adaptively to changes in the environment. Under ideal circumstances, the program’s behavior satisfies a set of application-dependent constraints. In the presence of failures, timing anomalies, or synchronization conflicts, however, certain constraints may become difficult or impossible to Satisfy, and the application designer may choose to relax them as long as the resulting behavior is sufficiently “close” to the preferred behavior. This paper describes the relaxation lattice method, a new approach to specifying graceful degradation for a large class of highly-concurrent fault-tolerant distributed programs. A relaxation lattice is a lattice of specifications parameterized by a set of constraints, where the stronger the set of constraints, the more restrictive the specification. While a program is able to satisfy its strongest set of constraints, it satisfies its preferred specification, but if changes to the environment force it to satisfy a weaker set, then it will permit additional “weakly consistent” computations which are undesired but tolerated. The use of relaxation lattices is illustrated by specifications for programs that tolerate (1) faults, such as site crashes and network partitions, (2) timing anomalies, such as attempting to read a value “too soon” after it was written, and (3) synchronization conflicts, such as choosing the oldest “unlocked” item from a queue.