Smart Routing at Scale: How Spotify's xDS Control Plane Cut 75% of Cross-Zone Traffic#

Yannick Epstein & Anya Hristova, Spotify

• Why: One of the few talks in the program with a hard, sustained production number attached to a specific architecture. Cross-zone traffic cost is a universal tax on multi-AZ deployment, and this quantifies what closing the control loop is worth.
• What: A sustained 75% reduction in cross-zone traffic across Spotify's compute infrastructure, stated as achieved without impacting reliability.
• How: They extended an in-house xDS control plane to do dynamic, data-driven zone-aware routing across a proxyless service mesh spanning two million nodes. The mechanism is Envoy's routing principles combined with real-time load telemetry: the control plane continuously recalculates optimal per-zone weights and pushes routing state in real time. The stated contrast is with their previous client-side routing, which optimized on each service's limited local view and so could not optimize globally.
• Where applicable: Spotify-specific in implementation — in-house control plane, proxyless mesh, two million nodes — but the speakers claim the design principles generalize to any large-scale mesh trading cost against availability, and the local-view-versus-global-view argument clearly does.