Continual learning improves Internet video streaming.

We describe Fugu, a continual learning algorithm for bitrate selection in streaming video. Each day, Fugu retrains a neural network from its experience in deployment over the prior week. The neural network predicts how long it would take to transfer each available version of the upcoming video chunks, given recent history and internal TCP statistics. We evaluate Fugu with Puffer, a public website we built that streams live TV using Fugu and existing algorithms. Over a nine-day period in January 2019, Puffer streamed 8,131 hours of video to 3,719 unique users. Compared with buffer-based control, MPC, RobustMPC, and Pensieve, Fugu performs better on several metrics: it stalls 5-13x less and has better and more stable video quality, and users who were randomly assigned to Fugu streamed for longer on average before quitting or reloading. We find that TCP statistics can aid bitrate selection, that congestion-control and bitrate-selection algorithms are best selected jointly, that probabilistic transfer-time estimates that consider chunk size outperform point estimates of throughput, and that both stalls and video quality have an influence on how long users choose to keep a video stream playing. Fugu's results suggest that continual learning of network algorithms in situ is a promising area of research. To support further investigation, we plan to operate Puffer for several years and will open it to researchers for evaluating new congestion-control and bitrate-selection approaches.