Video Streaming

A video platform is the purest read-heavy, bandwidth-bound system in the interview canon. The catalog is enormous and mostly cold, but at any moment a handful of clips are white-hot: a trailer drops, a clip goes viral, and a few hundred thousand people request the exact same four-second segment within the same minute. If those requests reach your origin servers and object store, both your tail latency and your bandwidth bill detonate. The entire engineering problem is to make sure they almost never do. This guide builds the design the SysGym simulator grades you against. The scenario offers a baseline of 20,000 segment requests per second that ramps to 4x during a viral surge, then injects three failures back to back: an origin node dies, the object store blips, and a CDN edge degrades. Your job is to keep playback succeeding at 99 percent or better, hold segment p99 under 80 ms, sustain at least 20,000 segment requests per second, keep the origin under 85 percent saturation, and do it all on a footprint of eight instances or fewer. Hit every one of those and you arrive, naturally, at the reference architecture: a high-hit-ratio CDN in front of an origin cache in front of a small, asynchronous origin, all reading immutable adaptive-bitrate chunks from object storage.