Improving L4 Routing Consistency
summary
Layer 4 load balancing misroutes when both a load balancer and a backend change because consistent hashing plus per-host LRU caches lose connection affinity. Two deployed fixes are described: embedding a QUIC connection ID server-ID hint (drives misroutes to near zero for QUIC) and adding a BPF TCP option carrying a server ID for internal TCP (adds ~6 bytes per packet, needs control of client and server). For public clients a shared flow cache broadcasts 5‑tuple→backend state for premium live streams and cuts misroutings by about 90%.
AI-written summary · about
chapters (8)
- 0:00L4 routing misroutings at Meta
- 1:53Layer four load balancers
- 5:24Failure scenarios in misroutings
- 7:28Meta’s load balancer history
- 9:56QUIC connection ID routing
- 13:04TCP option routing
- 17:25Shared flow cache
- 20:37Summary
related talks
46mKeeping the Balance: Load Balancing Demystified
17mQDSR: Accelerating Layer-7 Load Balancing by Direct Server Return with QUIC
34mHyperRouter: Lessons Learnt from Building an L4 Load...
30mKeeping the Balance: Internet-Scale Loadbalancing Demystified
19mA High-Speed Load-Balancer Design with Guaranteed Per-Connection-Consistency
23mBalancing on the Edge: Transport Affinity without Network State
32mLive Q+A 2
18mImproving Transfer Times in the Backbone Network Using QUIC Jump Start