Improving ABR at Netflix
summary
Netflix processes millions of playback sessions across devices and networks to measure where ABR decisions fail, finds that computing the true optimal per-session bitrate plan is NP-hard (equivalent to the multi-choice multi-period knapsack), and adopts a linear-time approximation that prioritizes bandwidth to earlier chunks because of tighter deadlines. The team ships a scalable API review assistant in production that compares observed throughput and encoding choices to approximated optimums and surfaces the largest practical wins.
AI-written summary · about
related talks
16mImproving Streaming Experience with Bayesian Optimization, from AB to...
22mProviding better video experiences for the next billion users.
19mUsing application informed pacing to be a friendly internet neighbor
17mDelivering better video encodes for legacy devices
13mUltra-High Resolution Playback Visualization and its Applications
19mReinforcement learning for ABR
10mGoing Beyond Bitrate: Adaptation Based on Video Context
20mStreaming at 250 kbps: raise your expectations