Distributed Consensus Algorithms
summary
Distributed consensus is what prevents split-brain problem and lets services like Photon, Chubby, ZooKeeper and Spanner give strong consistency by ordering operations with Paxos or Raft. Production deployments typically use five replicas -> majority of three for fault tolerance; Paxos requires two rounds but multipaxos 50% speed-up collapses that to one during normal leadership. Global deployments hit latency limits — a ~100 ms transatlantic latency bounds commits to ~10/sec unless batching, pipelining or hierarchical quorums are used.
AI-written summary · about
chapters (11)
- 0:00Start
- 13:46Distributed consensus problems
- 14:29ACID vs BASE
- 15:29Why ACID?
- 16:54Why NOT strong consistency?
- 18:48Distributed Consensus: a brief history
- 25:26Other consensus algorithms
- 34:53Scaling throughput
- 41:47Scaling read throughput
- 44:36Monitoring
- 46:15Further Reading
related talks
40mConsensus Algorithms in Distributed Systems
42mConsensus Algorithms in Distributed Systems (Theory and Practice)
43mAll Together Now: An Introduction to Distributed Consensus
38mTowards consensus on distributed consensus #bbuzz
51mDo Not Reimplement Distributed Consensus Algorithms
23mLightning Talks on Distributed Systems
22mLiberating Distributed Consensus
40mImplementing Distributed Consensus