Netflix Infrastructure supports personalized UI and Streaming experience across 230M+ members around the world. Spread across multiple locations, it’s important to have predictability and control over how user traffic is balanced across them, to ensure balance between latency, infrastructure costs, and availability risk. 

This talk will tell a story of how Netflix has shifted from geo-based DNS load-balancing to latency-based approach, relying on real-user measurements and building a global data model of Netflix traffic to reduce costs while reducing latency and outage risks. We will cover challenges of integrating the solution into Cloud and CDN components of Netflix infrastructure, and trade-offs between model accuracy and traffic model complexity. The talk also demonstrates how the data-driven approach was applied to influence future infrastructure decisions, by simulating impact of potential infrastructure changes with precision and minimal engineering effort.