An Uncoordinated Approach To Scaling (Almost) Infinitely
summary
Argues that systems scale near linearly by avoiding cross-node coordination: partition state into entities that fit on one machine, process incoming messages serially so each entity maintains local serializability, let business logic return outgoing messages while a scale-aware layer routes them, and treat inter-entity links as activities managed by the platform. Uses at-least/at-most-once delivery tradeoffs and duplicate handling or CRDT-style merges; relies on I-confluence to prove when coordination can be safely omitted.
AI-written summary · about
chapters (8)
- 0:00Start
- 0:52Andrew Turley
- 1:33The Big Idea
- 2:15What to Expect Here
- 2:59Overview
- 11:20Life beyond Distributed Transactions
- 26:40Coordination Avoidance in Database Systems
- 32:05Conclusion
related talks
46mPerfect Scalability
38mHow Scale Makes Distributed Systems Slower
44mDistributed systems and scalability rules
28mPractical Approaches To Scaling Distributed Transactions
14mCreating tools to cross machines and networks
19mEngineering for the Long Game
42mLarge-Scale Architecture: The Unreasonable Effectiveness of Simplicity
25mThe shell game called eventual consistency