As Kubernetes clusters grow beyond hundreds of microservices, iptables routing rules degrade O(N) into CPU bottlenecks. By shifting connection state tracking directly into the Linux kernel with eBPF, we achieved sub-millisecond network policy verification.
The Collapse of kube-proxy at Scale
Every Kubernetes service creates iptables entries that the kernel traverses linearly on every packet. At 5,000 services, sequential rule evaluation introduces measurable packet jitter and saturates softirq cycles.
Cilium replaces iptables by attaching eBPF bytecode programs to the Traffic Control (tc) and socket layers of the Linux kernel, routing packets using constant-time BPF hash maps (O(1)).




