The Road to Akka Cluster, and Beyond
summary
Grounding distributed theory (FLP, CAP, Lamport clocks) into an implemented system, Akka Cluster uses epidemic gossip with vector clocks, a SWIM-like failure detector and CRDT work to provide decentralized membership, cluster sharding, singleton and pub/sub. It claims pragmatic tradeoffs (favoring availability with fallbacks to consistency), changes autodown to manual after split‑brain issues, and demonstrates scale — booting 1,000 nodes in four minutes. Raft and ACID 2.0 experiments are underway for stronger consistency.
AI-written summary · about
chapters (26)
- 0:00Intro
- 3:15WHY DO WE NEED IT?
- 4:13YOU CAN'T TELL THE DIFFERENCE
- 4:53FALLACIES
- 5:26A MODEL FOR DISTRIBUTED COMPUTATION
- 6:15Lambda Calculus (Alonzo Church, 1930)
- 7:33VON NEUMANN MACHINE JOHN VON NEUMANN 1945
- 8:53Actors (Carl Hewitt, 1973)
- 11:37IMPOSSIBILITY OF DISTRIBUTED CONSENSUS WITH ONE FAULTY PROCESS
- 14:14LINEARIZABILITY
- 15:11DISSECTING CAP
- 17:15LAMPORT CLOCKS
- 18:28VECTOR CLOCKS
- 20:19Failure Detection: Formal Model
- 21:31ACCRUAL FAILURE DETECTOR HAYASHIBARA ET AL. 2004
- 23:09SWIM FAILURE DETECTOR
- 25:52SPECIFICATION
- 29:30NODE RING & EPIDEMIC GOSSIP
- 33:27Akka CLUSTER EXTENSIONS
- 34:43CLUSTER MEMBERSHIP IN AKKA
- 35:27GOSSIPING
- 38:42LEADER
- 40:44NETWORK PARTITIONS
- 42:14AKKA MODULES FOR DISTRIBUTION
- 42:47AKKA & THE ROAD AHEAD
- 43:55REFERENCES
related talks
46mBuilding high available systems with Akka Distributed Data
55mExperience report on using Akka in production: pitfalls and rewards
41mAkka in Production: Our Story
48mThe three paths of Scaling with Akka
56mAkka + F# = Akkling
45m(...)Distributed Systems
52mThe best is yet to come - State of Akka in 2017
57mScaling the CAP Theorem for Modern Distributed Systems