A Knowledge Theoretic Account of Recovery in Distributed Systems: The Case of Negotiated Commitment

We are interested in commitment problems in potentially faulty distributed environments; for such problems, the behaviour of failed processes during recovery is relevant to consistency. In particular, we examine negotiated commitment, which is the problem of ensuring that each participant in a negotiation reaches a consistent local decision on the outcome. Even undecided, recovering participants must reach a consistent decision on the outcome, because other participants may have committed to an outcome and taken further actions based upon the expected commitment of the recovering participant. To facilitate the use of knowledge theory to guide the design of protocols for commitment problems, we give an account of process failure and recovery. Using knowledge theory, we show that independent recovery is impossible --- i.e., a recovering participant whose decision must be based on some knowledge about other participants in the system cannot decide upon recovering without communicating with other participants. If the participant is in a decided state upon recovery without such communication, then it must have been decided when it failed, and furthermore it must have been decided before it failed. We also give levels of interparticipant knowledge necessary for achieving nonblocking recovery in the absence of total participant failure.