Consensus Algorithms in Distributed Systems
summary
Explains how Paxos: prepare/accept with proposal numbers and Raft: leader election + log replication actually operate, including Paxos proposers/acceptors, proposal-number construction, Raft election timeouts (typical 100–300 ms), majority rule (floor(n/2)+1), AppendEntries/heartbeat mechanics and log-matching repair steps. Cites industry uses (Google Chubby, ZooKeeper, Kafka relying on centralized consensus), reports only ~13% of surveyed devs can describe an algorithm, and warns Byzantine tolerance is possible but far more expensive.
AI-written summary · about
chapters (20)
- 0:00Introduction and setup
- 1:16Why talk about cakes first
- 3:58Distributed systems challenges
- 7:11Consensus algorithms overview
- 7:34Paxos explained with cake analogy
- 10:04Paxos step-by-step with cake vote
- 14:16Multi-Paxos and leader role
- 16:02Industry examples of Paxos
- 16:23Raft introduction and simplicity
- 17:48Raft leader election with cake vote
- 18:56Raft leader election phases
- 21:24Ensuring leader election in Raft
- 23:10Raft log replication process
- 25:59Detecting and fixing log inconsistencies
- 27:04Voting criteria in Raft
- 30:36Repairing log inconsistencies
- 33:48Log matching property in Raft
- 34:10Problems with consensus algorithms
- 37:16Takeaways and implementation advice
- 39:26Q&A session
related talks
42mConsensus Algorithms in Distributed Systems (Theory and Practice)
48mDistributed Consensus Algorithms
51mDo Not Reimplement Distributed Consensus Algorithms
43mAll Together Now: An Introduction to Distributed Consensus
07mA Walk Through Distributed Consensus
01h 13mPaxos vs Raft: Have we reached consensus on distributed consensus?
47mDistributed Systems: When Democracy and Computer Science Come Together
23mLightning Talks on Distributed Systems