Techniques Netflix Uses to Weather Significant Demand Shifts
summary
Netflix runs a four-region full-active control plane backed by Apache Cassandra, EVCache and Zuul and steers traffic via ~100 authoritative DNS points on Open Connect; weekly failover drills move a region’s traffic in minutes. Services use predictive prescaling plus tens of thousands of EC2 autoscaling groups with RPS hammers to cut recovery from ~8–15 minutes to ~3–4 minutes. Load protection uses prioritized load shedding (client priorities 0–100), data gateways that chunk/paginate and add idempotency tokens, memcache fronting, and Hollow snapshots for in-memory read scale.
AI-written summary · about
chapters (17)
- 0:00Intro
- 0:40Netflixs Problem
- 5:05Global Traffic Demand
- 11:30Service Headroom
- 19:18Prescale
- 26:16Load shedding
- 29:44Highway Analogy
- 30:24IO workloads
- 31:48Retries
- 33:51Queue Theory
- 34:39Data Gateway
- 37:49Real World Example
- 38:35Abstraction Series
- 39:00Caches for Availability
- 40:44Hollow
- 41:44Test
- 42:55Conclusion
related talks
21mHandling Traffic Spikes Through Automated Prioritized Load Shedding
39mAvailability, Latency, and Cost: Withstanding Regional Outages
24mHow Netflix enables
33mAvailability, Latency and Cost: Withstanding Regional Outages
20mBetter Resilience Through Self Tuning
19mKeynote: Making and Keeping Netflix Highly Available
27mBehind the Scenes: Running Critical Services @ Netflix
46mCassandra Serving Netflix @ Scale