Consistency without consensus in production systems
summary
Argues that provable consensus‑free consistency is practical: CALM, ACID 2.0 and CRDTs yield real guarantees; SoundCloud builds Roshi, a production CRDT-backed timeline store implemented in ~2,000 lines of Go that uses Redis sorted sets as primitives, sharding by outbox key, replicated clusters, a farm coordinator, background read‑repair and an eight‑hour walker to converge state. Roshi enables fan‑in‑on‑read timelines, tolerates duplicates and reordering, and runs in production serving high‑volume time‑series.
AI-written summary · about
chapters (9)
- 0:00Start
- 0:51Distributed programming
- 2:37Idioms
- 12:15Increment-only counter
- 18:55Interlude — Bending the problem
- 27:42A wild set appears
- 29:00Reading is easy
- 30:02Example
- 35:19Reading is interesting
related talks
57mLiving in eventually consistent reality
46mCoordination-Free Computations
36mConflict-Free Replicated Datatypes (CRDT)
48mThe Accord Consensus Protocol, how to do ACID transactions at scale
50mCRDTs in Practice
42mBeyond Blockchain: Convergent Consensus
14mBolt-On Strong Consistency: Specification, Implementation, and Verification
25mThe shell game called eventual consistency