A middleware for gossip protocols

Gossip protocols are known to be highly robust in scenarios with high churn, but if the data that is being gossiped becomes corrupted, a protocol’s very robustness can make it hard to fix the problem. All participants need to be taken down, any disk-based data needs to be scrubbed, the cause of the corruption needs to be fixed, and only then can participants be restarted. If even a single participant is skipped in this process, say because it was temporarily unreachable, then it can contaminate the entire system all over again. We describe the design and implementation of a new middleware for gossip protocols that addresses this problem. Our middleware offers the ability to update code dynamically and provides a small resilient core that allows updating code that has failed catastrophically. Our initial PlanetLab-based deployment demonstrates that the middleware is efficient.