Long-Lived User-Session VM: Enhancing Zero Downtime Rolling Deployments for Virtual Mac...
summary
GeForce Now uses a per-instance keepalived VRRP standby and a Kubernetes operator CR reconciliation to roll reverse‑proxy VMs with long TCP sessions, syncing iptables and session maps over HTTP to a standby, promoting it by bumping keepalived priority, then deleting the old VM after a one minute safety window. Drain time for sessions can be eight hours, rollout time drops from days to hours, and clusters run on the order of 1–2k VMs per cluster.
AI-written summary · about
chapters (13)
- 0:00Introduction
- 0:20Agenda
- 0:56What is GeForce Now
- 4:51Challenges
- 6:57Solution Overview
- 9:00What is VRRP
- 11:20Highlights
- 12:30KeepLiveD
- 13:22State Synchronization
- 14:35Reversal
- 15:17Repeat
- 16:01Conclusion
- 20:50Questions
related talks
43mHot Deploying Low-Latency Services for 24/7 Operation
14mDeep Optimization of VMM Live Upgrade
37mZero-downtime Deployment Strategies (...) [ENG]
37mZero downtime Deployment on Kubernetes
15mKeeping a Hypervisor Fleet Up to Date with Minimal Customer Disruption
33mThe Gotchas of Zero-Downtime Traffic /w Kubernetes
15mAre You Vmweary?
40mStrategies for Zero Down Time, Frequent Deployments