Balancing on the Edge: Transport Affinity without Network State
summary
Faild is a distributed load balancer deployed at Fastly since 2013 that uses commodity switches plus a tiny host kernel module to achieve stateless operation and graceful failover by encoding current and previous target hosts into virtual MACs. Switches do bulk forwarding; if a packet hits a remapped entry the kernel redirects it to the original server, adding a median detour of about 14 microseconds and under 0.3% CPU overhead. Operational scale claims: half-rack ~320k RPS, largest PoP 4.8 Tbps / 2.5M RPS, and documented surprising vendor hashing bugs (e.g. 6× bucket imbalance, 255 vs 256 next-hop gaps).
AI-written summary · about
chapters (9)
- 0:00Start
- 4:38Requirements
- 5:33Problem statement
- 5:52Guidelines
- 7:29load balancing
- 11:08Host processing
- 14:25Uneven hashing
- 15:33Hash polarization
- 16:54Nope, things break
related talks
26mF10: A Fault-Tolerant Engineered Network
04mLoad balancing through the years
32mCDN Scaling Challenges
15mFail Open, Fail Fast: Improving Envoy Resilience in Latency-Critical Systems
46mKeeping the Balance: Load Balancing Demystified
17mBackdraft: a Lossless Virtual Switch that Prevents the Slow Receiver Problem
16mA High-Speed Stateful Packet Processing Approach for Tbps Programmable Switches
30mKeeping the Balance: Internet-Scale Loadbalancing Demystified