Reinforcement Learning-Driven Cache Refill for Live Stream Outage Recovery

Abstract

Live video streaming services must keep playing without stopping, even if the network goes down, which can cause viewers to rebuffer and leave (Mux Data, 2021). Standard edge-cache fill- ing methods, such as getting all the missed segments or going straight to live, have fixed trade-offs between content completeness and latency. However, they can’t adjust to different outage lengths and network conditions (S. Wang et al., 2018). We introduce RL-CacheRefill in this paper. It is a cache refill framework based on reinforcement learning that treats post-outage recovery as a Markov decision process (Boyan & Littman, 1994). The agent watches real-time data like buffer occupancy, outstanding segment gap, and backhaul throughput. It then learns the best way to decide whether to fetch or skip each missing segment. This method strikes a balance between the conflicting goals of keeping as much content as possible while minimizing stall time and playback lag (Mao et al., 2017). We use publicly available network traces and realistic live-stream traffic to create a sim- ulation environment that can be reproduced. We then compare RL-CacheRefill against full-fetch, skip-to-live, and threshold-based baselines. The results show that our solution cuts average rebuffer- ing time by up to 30% and playback latency by up to 25%. This means that the quality of experience is more than 20% better than with static policies. We look at how policies work in different network settings and give tips on how to use them in edge-cache servers. An AWS outage caused Netflix’s video service to be down for more than two hours in May 2011. This led to an estimated 2 million minutes of lost viewing time and a 20% rise in customer support tickets (Netflix Technology Blog, 2011). Mux says that if there is even a few seconds of buffering, 15–25% of people watching live will leave. So, recovering from an outage isn’t something that happens very often; it has a direct effect on retention and revenue.