Lightning Talk: Envoy @ Netflix
Lightningat EnvoyCon North America 2025 · on Monday, 2025-11-10· at 12:15· Building B | Level 4 | B406b-407·
summary
Envoy is deployed as sidecars on VMs, containers and laptops using ODCDS to resolve per-request cluster config and hot restart to let apps program ingress behavior; balancing issues surface as single-worker limits (about 1k vs 100 RPS per worker) and massive fan-out creates mnc connections reduced from ~100 million to ~50 million by connection subsetting. Key fixes: increase connection cardinality for spikes, pre-warm pools, make WASM VMs restartable, and drop Huffman to save ~5–10% CPU.
AI-written summary · about
chapters (11)
- 0:00Envoy at Netflix
- 0:43Why Netflix chose Envoy
- 1:48Incremental adoption and hot restart
- 2:51Too few connections between app and egress
- 4:13Too many connections in egress
- 5:42Not enough connections during spikes
- 7:30Filter chain isolation and shared resources
- 8:14Removing Huffman encoding for local
- 8:35Optimizing Envoy’s stat system
- 9:40Slow control plane with ODCDS
- 10:02Summary
related talks
33mGet to know Envoy - the modern proxy for cloud-native infrastructure - Baptiste Collard
34mIntro: Envoy
37mEnvoy Internals Deep Dive
09mLightning Talk: Protecting Envoy: Overload Manager
05mLightning Talk: Optimizing Cluster Workloads: Cilium Envoy on DPU
31mEnvoy Project Intro – Matt Klein & Jose Nino, Lyft
38mEnvoy Deep Dive
21mLightning Talk: Integrating Envoy with DPDK-based virtual networks