How Netflix Builds High Performance Applications at Global Scale

We all want to build applications that are blazingly fast. We also want to scale them to users all over the world. Can the two happen together? Can users in the slowest of environments also get a fast experience? Learn how we do this at Netflix: how we understand every user's needs and preferences and build high performance applications that work for every user, every time.

1. How Netflix Builds High Performance Applications at Global Scale Prasanna Vijayanathan Senior Software Engineer at Netflix

2. Prasanna Vijayanathan ( he/him ) Senior Software Engineer at Netflix Something cool I’ve done My perspective on P99s Another thing about me What I do away from work

3. Your app (phone, browser, TV, multiverse?)

4. Your app (phone, browser, TV, multiverse?)

5. Your infrastructure Your app (phone, browser, TV, multiverse?) Reverse proxy/load balancer Data Microservices

6. Your infrastructure Your app (phone, browser, TV, multiverse?) Reverse proxy/load balancer Data Microservices CDNs

7. Your infrastructure Your app (phone, browser, TV, multiverse?) Reverse proxy/load balancer Data Microservices CDNs Users

8. Understanding your application Observability and monitoring

9. Your infrastructure Your app (phone, browser, TV, multiverse?) Reverse proxy/load balancer Data Microservices CDNs Users Errors, latency

10. Your infrastructure Your app (phone, browser, TV, multiverse?) Reverse proxy/load balancer Data Microservices CDNs Users Errors, latency Tracking

11. Your infrastructure Your app (phone, browser, TV, multiverse?) Reverse proxy/load balancer Data Microservices CDNs Users Errors, latency Tracking

12. Your infrastructure Your app (phone, browser, TV, multiverse?) Reverse proxy/load balancer Data Microservices CDNs Users Errors, latency Tracking

13. Your infrastructure Your app (phone, browser, TV, multiverse?) Reverse proxy/load balancer Data Microservices CDNs Users Errors, latency Tracking

14. Your infrastructure Your app (phone, browser, TV, multiverse?) Reverse proxy/load balancer Data Microservices CDNs Users Errors, latency Tracking

15. User Experience

16. Your infrastructure Your app (phone, browser, TV, multiverse?) Reverse proxy/load balancer Data Microservices CDNs Users

17. Your app (phone, browser, TV, multiverse?) Users Engagement Sessions Time spent on app Environment Platform Device capabilities Network conditions Experience Time to load the app Time to render Video quality

18. App load time App load time

19. Diversity at scale

20. Diversity at scale Platform Memory Computing & efficiency Device Location Network Geo Needs Preferences Users

21. Diversity in devices Web Mobile TV Other devices (VR, Watches, Glasses,..) Platform Crashes and errors Storage Buffering Memory CPU/GPU Energy consumption Computing and efficiency

22. Diversity in network Ethernet, Fiber, WiFi, 4G, 5G Type of Network 1000s of providers different plans ISPs bandwidth, throughput, errors Factors impacting networks

23. Diversity in Users What does the user need from the app? User's needs How does the user expect the app to behave? User's preferences

24. Consistency of QoE

25. Consistency of QoE Experience is relative Users like to get expected behavior

26. Consistency of QoE Provide the expected behavior every time Do better whenever possible

27. Consistency of QoE How to know what’s expected? User disruptions

28. User Disruptions Needs & Preferences Environment

29. Needs & Preferences App improvements New product offerings, features, performance Better feedback Lower support calls Longer sessions, more engagement User Needs & Preferences Distruptions Environment

30. Disruptions Errors, crashes Action fails Loading spinners User Disruptions Needs & Preferences Environment

31. Environment User Environment Needs & Preferences Device Make/model, platform, memory, CPU Network throughput, latency, errors Others time, location, seasonality Disruptions

32. User Disruptions Needs & Preferences Environment

33. How could we do this? QOE models Devices Networks Disruption prediction Transfer learning Learn what’s good and bad for each user

34. How could we do this? <Demo: building a model with a sample data set>

35. How could we do this? <Demo: Show using SHAP values for interpretability>

36. What’s possible? Constant iteration and improvement Automatic feature selection and experimentation The best experience possible, every user, every time.

37. First Last [email protected] @twitter website/blog url Thank you! Let’s connect.