Measuring Real-Life Latency of the Internet: A Netflix Story
summary
Netflix measures per-session round-trip-time distributions using t-digest so it can store and merge summaries at scale instead of raw samples. The CDN serves 230 million members from ~18,000 servers with a 14-person reliability team, so low-overhead stats matter. Typical session metrics show p50 ~115 ms, P99 ~2 s while a max can reach ~4 s, and these summaries detect anomalies (e.g., during the March 4 Chris Rock live broadcast) and drive capacity placement and routing fixes.
AI-written summary · about
chapters (7)
- 0:00Personal introduction
- 2:01Netflix infrastructure
- 4:20Measuring user experience
- 5:45Round-trip time data
- 7:37Using T-Digest
- 9:25Extracting insights
- 11:08Monitoring and conclusion
related talks
22mLancet: A self-correcting Latency Measuring Tool
10mMeasuring the live video latency at scale
31mMastering Near-Real-Time Telemetry and Big Data: Invaluable Superpowers for Ordinary SREs
32mSoftware Performance: A Shape Not a Number
37mLies, Damn Lies, and Metrics
38mTechniques and Tools for a Coherent Discussion about Performance in Complex...
28mIntroduction to WebRTC Statistics and getStats() API
17mMetrics Driven Obsession with Viewer Experience