How would you architect a feature-store pipeline to serve model predictions with sub-100ms latency on Twitch?
