• Sources: primary
  • Summary: Pods in a customer's europe-west GKE clusters timed out reaching the kube-apiserver after a February 2026 Private Service Connect switch update began dropping RFC 5961 challenge ACKs, which NAT port reuse triggered often. Google rolled that change back, then found PSC packet reordering, Envoy half-closed sockets that tcp_fin_timeout never reaps, and a kernel 6.1 tcp_ack() path that ignored a reordered FIN. The final fix landed as a Linux kernel patch.
  • Why it matters: Proxies that half-close sockets behind NAT can leave FIN_WAIT_2 state that collides with reused ports, a failure mode worth checking in Envoy-fronted services.