Are you Dead or Alive? Using Gossip for Failure Detection
summary
Gossip protocols claim fast, scalable failure detection by having each node gossip only to a few peers so updates diffuse in log n rounds; Xerox’s 1987 epidemic paper shows replication across ~300 nodes and practical systems (Cassandra, Dynamo, Amazon S3, Hashicorp’s surf) adopt variants. Implementations use anti-entropy exchanges or rumor mongering with push/pull, a tunable K parameter, and provide eventual consistency and fixed per-node messaging; tradeoffs are redundant bandwidth and latency and failure when injection rate exceeds gossip capacity.
AI-written summary · about
chapters (14)
- 0:00Start
- 1:35Takeaways
- 14:01Motivations
- 16:03Types of Epidemics
- 16:22Direct Mail
- 18:07Anti-Entropy
- 24:48Gossip/Rumor Mongering
- 29:00Messaging Strategies
- 29:52Uses of Gossip
- 31:52Where is Gossip used
- 33:59Strengths of Gossip
- 36:17Weaknesses of Gossip
- 37:39Variations and Other Algorithms
- 41:31References
related talks
45mLet's talk Gossip!
46mBuilding high available systems with Akka Distributed Data
44mThe Road to Akka Cluster, and Beyond
40mAll In With Determinism for Performance and Testing in Distributed Systems
43mDistributed Systems
45mConcurrency + Distribution = Scalability + Availability (...)
45m(...)Distributed Systems
38mHow To Build a Distributed System (And Should You?)