Inside Spotify’s Envoy Architecture: What We Le...
Talkat EnvoyCon Europe 2026 · on Wednesday, 2026-03-25· at 11:10· Hall 1-5 | Tram Zone | Cloud Native Theater·
summary
Spotify runs Envoy at an 11 million requests per second scale and reports a real improvement from ~500 to ~2,000 rps per core (later 4,500) after migrating fully to Kubernetes and replacing a complex Java sidecar with Envoy-native C++ filters, while standardizing on gRPC. Migration uses side-by-side duplicated requests to compare behavior, kept token prefetch as an init container, and hit a production outage caused by changed filter order that triggered OOMs, showing concrete latency, reliability, and cost trade-offs.
AI-written summary · about
chapters (11)
- 0:00Envoy at Spotify
- 1:02Envoy’s infrastructure role
- 2:26Java sidecar challenges
- 3:51Migration plan for Envoy
- 5:15Side-by-side comparison
- 6:40Deployment issues
- 8:06Filter order incident
- 9:12Migration lessons
- 10:19Performance improvements
- 11:25Summary
- 11:46Q&A
related talks
34mThe Spotify Story: Moving from NIH to an Active CNCF Community Member
33mGet to know Envoy - the modern proxy for cloud-native infrastructure - Baptiste Collard
22mUnlock Kubernetes & Envoy: Master Your API Ecosystem!
35mTo Infinite Scale and Beyond: Operating Kubernetes Past the Steady State
28mDeploying hybrid topologies with k8s & Envoy: a look at service discovery
36mReplacing NGINX with Envoy in a Traffic Control System
17mLessons Learned: Four Years with Emissary-ingress and Envoy
30mSpotify Harnesses Ray for Next-Gen AI Infrastructure