gRPC Proxyless Service Mesh
summary
Spotify migrates from Hermes/Nameless/Helios toward a gRPC proxyless service mesh using Google Traffic Director; Hermes is a homegrown protobuf/ZeroMQ protocol that lacks zone-aware load balancing, Nameless is DNS-based and hits a 64 KB response limit (~800–900 hosts) which breaks at scale, and Helios is being replaced by Kubernetes. A 2022 rollout causes a global outage, so Google adds canary rollouts and improved observability. Migration requires per-service protocol changes, incentives, and careful rollout.
AI-written summary · about
chapters (14)
- 0:00Start
- 1:49A Painful migration
- 2:40Migrating herds and herding migrations
- 4:19The fine technology migration balance
- 6:30Hermes - a protocol
- 7:11Nameless - a service discovery service
- 8:25Helios - a container platform
- 9:30But DIY comes with overhead
- 11:10Why no Nameless
- 12:24The cross-zone traffic problem
- 15:33The dream - proxyless service mesh
- 16:55Why proxyless
- 17:29So what is stopping us?
- 18:20The Traffic Director incident
related talks
30mOops!...I Did It Again — I Reinvented gRPC
29mThe Story of Why We Migrate to gRPC and How We Go About It
31mHow We Moved Spotify To a Proxyless gRPC Service Mesh
36mAdopting gRPC at Spotify
35mThe Story of why we migrate to gRPC and how we go about it
23mA Simplified Service Mesh With gRPC
22mUsing a Service Mesh With gRPC
18mHigh Performance Service Meshes with gRPC Proxyless