distributed-consensus-with-raft.md

DISTRIBUTED CONSENSUS WITH RAFT

@jxxf

• Why consensus is hard
• Distributed consensus
• Many ways, many implementations
• 2 general buckets:
  • Symmetric (everyone agrees)
  • Asymmetric (designated leader)
• Locality bubble
• Two generals problem A, B
  • both want attack a city
  • city can repel 1 but not both attacks
  • generals should co-ordinate their attacks to ensure destruction of city
  • city has spies to intercept communication
  • city can only transform, not block messages
  • impossible to coordinate, cannot confirm receipt of message
• Distributed systems are hard, yo
  • Blockchains are one implementation, mostly really hard Math problems
  • Paxos, but hard to implement, maintain
    • Goal of Raft: available, replicated state machine
    • Raft is asymmetric
      • Leader, Candidate, Follower
      • Everyone starts as a Follower
      • After a timeout, becomes a Candidate
      • If enough votes, Candidate becomes Leader
      • If not enough votes, timeout again, start over
      • Election valid outcomes: 3
• Raft: Used by Kubernetes, Docker Swarm, and more
• …

---

Metadata

Tue Jun 28 14:22:28 PDT 2016