Do Not Reimplement Distributed Consensus Algorithms
summary
Compares two-phase commit, Paxos and Raft with concrete failure scenarios, shows why two-phase commit blocks on a single coordinator, explains Paxos ballots and how majority quorum recovers progress, and shows Raft’s leader-based replication, election timeouts and term rules. Demonstrates Cassandra’s Paxos-backed lightweight transactions cost four round trips and warns not to mix LWT with normal writes; covers Mongo’s Raft settings (ack counts, journaling) and the practical trade-offs between safety, liveness and performance.
AI-written summary · about
chapters (16)
- 0:00Introduction and setup
- 1:08Speaker introductions
- 3:15Why consensus is hard
- 7:29Two-phase commit downsides
- 8:56Paxos overview and properties
- 12:51Paxos failure scenarios
- 17:30Paxos learners and roles
- 22:45Paxos leader election
- 23:27Paxos in Cassandra
- 29:53Raft introduction
- 35:10Raft log replication
- 40:10Raft leader failure handling
- 44:50Raft properties and guarantees
- 45:34Raft in MongoDB
- 47:19Raft configuration trade-offs
- 49:06Applying consensus knowledge
related talks
40mConsensus Algorithms in Distributed Systems
47mDistributed Systems: When Democracy and Computer Science Come Together
01h 13mPaxos vs Raft: Have we reached consensus on distributed consensus?
43mAll Together Now: An Introduction to Distributed Consensus
47mUnderstanding Raft Consensus
48mDistributed Consensus Algorithms
43mWhat We Talk About When We Talk About Distributed Systems
42mConsensus Algorithms in Distributed Systems (Theory and Practice)