He traced the config history. Turned out, a junior engineer had, as a joke 14 months earlier, set a max_ttl_days=0 in a feature flag config — meaning "no timeout." But the flag parser had a bug: 0 got stored as nil , and nil in their system defaulted to . The VM was literally older than the region’s deployment pipeline version .
Alex and his team spent 11 hours patching the VM config parser, manually draining the zombie VM, and replaying 14 months of missing model snapshots. Post‑mortem title: “A VM walked into a bar and never left.” netflix vm config
Then came the really weird part. Because the VM never recycled, its local SSD (ephemeral) had accumulated — normally deleted every week. The ML training pipeline saw this "ancient" VM as a stable node and started preferring it for critical A/B tests. By December 23rd, 3% of all北美 traffic was being routed through this single zombie VM. He traced the config history
Alex dug into the VM’s birth certificate (a metadata endpoint they used for auditing). The VM was provisioned — impossible, because Netflix autoscaling recycled VMs every 14 days max. Alex and his team spent 11 hours patching