SLATE: Intelligent Request Routing
summary
SLATE is a Service‑Layer Traffic Engineering system that globally optimizes request routing using telemetry from an Envoy WASM plugin, cluster controllers and a global controller that runs a linear regression load→latency model and a linear‑programming optimizer. It trades off latency versus egress cost with an admin intent metric like dollars per millisecond, prioritizes request classes, and in a demo shifts traffic across two clusters to cut average latency from ~300 ms to 52 ms avg / 176 ms tail.
AI-written summary · about
chapters (16)
- 0:00SLATE project intro
- 1:23Routing challenges in Microsoft apps
- 2:49Cross-cluster routing gaps
- 4:14Traffic management limitations
- 4:56SLATE design from traffic engineering
- 5:37SLATE architecture and data flow
- 7:44Latency optimization under load
- 9:50Comparison with baseline routing
- 10:33Dollar per millisecond metric
- 11:36Minimizing egress costs
- 13:22Handling request classes
- 14:24Challenges and future plans
- 15:05Demo: latency optimization
- 18:37Q&A: control flow and API
- 19:59Q&A: local vs global decisions
- 20:43Q&A: DNS and high availability
related talks
25mResQ: Enabling SLOs in Network Function Virtualization
16mExpress Lane to Efficiency and Reliability: Multi-Dimensional Control in Meta’s Express..
29mSpanning the Globe with Envoy at Stripe
35mIntelligent Load Balancing in Kubernetes
37mLevel Up Your Edge: Do You Know Where Your Traffic Comes From?
19mDecentralized cloud wide-area network traffic engineering with BLASTSHIELD
23mNetwork SLOs: Knowing When The Network Is The Barrier To Application Quality Of Experience
21mR2P2: Making RPCs first-class datacenter citizens
