The Evolution of Traffic Routing in a Streaming World
summary
A team migrates from co‑located services to an off‑box fleet and then to a full‑mesh where every gateway connects to every Blade Runner. They solve connection limits by using rendezvous hashing so each gateway picks k = 3 nexuses with minimal churn. Full mesh enables sticky routing per feed, reduces KV reads by 15–20%, saves about 5% memory and 7–10% CPU, and uses virtual pools to auto‑scale a customer from a handful of hosts to many during a spike without affecting others.
AI-written summary · about
chapters (11)
- 0:00Introduction and context
- 1:04Co-located architecture
- 2:26Off-box architecture
- 3:48Rendezvous hashing basics
- 5:28Solving routing challenges
- 8:02Dynamic load balancing
- 9:27Full mesh architecture
- 11:23Sticky routing implementation
- 13:14Virtual isolation pools
- 15:18Scaling and pool management
- 19:13Summary and key takeaways
related talks
17mMeta’s Network Journey to Enable AI
33mCheap Routes, Fast Paths & Keeping Track
28mGateway Architecture 3.0
48mRethinking Connectivity
13mEfficient Navigation for Modern Networks
29mLessons Learned in the Move to IP Routing
16mExpress Lane to Efficiency and Reliability: Multi-Dimensional Control in Meta’s Express..
21mServiceRouter: Hyperscale Service Mesh