Availability, Latency, and Cost: Withstanding Regional Outages
summary
Netflix moves from an island model to a 3-region (2+1) N+1 model, cutting failover time from ~45 minutes to 7 minutes with Project Nimble and running evacuations biweekly so failovers are routine. Capacity math shows three regions cost ~150% of peak vs 200% for two, and cost depends on a database portion vs throughput. Engineering choices include replicate caches, avoid cross-region calls, and two-copy sharding patterns (central-savior or ring) to halve data replication cost while preserving availability.
AI-written summary · about
chapters (20)
- 0:00Introduction
- 0:34Agenda
- 1:14Netflix
- 2:17Island Model
- 3:23Fear
- 6:56Recovery vs Prevention
- 8:18High Availability Overview
- 11:53Algebraic Models
- 14:01Latency
- 15:12Cumulative Distribution Function
- 16:28Cost
- 20:32Architecture
- 21:49Sharding
- 24:23Region Replication
- 27:47Closing Thoughts
- 29:08Questions
- 29:58Summary
- 32:58Cost Function
- 35:08Model Change
- 36:09Alternatives
related talks
44mTechniques Netflix Uses to Weather Significant Demand Shifts
24mHow Netflix enables
39mCloud Native Architectures
31mKeeping Your Cloud Footprint in Check
48mResiliency and Availability Design Patterns for the Cloud
19mKeynote: Making and Keeping Netflix Highly Available
26mBuilding Financial Systems on Eventually Consistent DBs
46mCassandra Serving Netflix @ Scale