article --summary
The summary
The post explains that the ipvs kube-proxy backend, once added to fix iptables performance problems, never fully matched Kubernetes Service behavior and is now being retired. Kube-proxy's ipvs mode depends on iptables for some features and lacks batch updates, which limits its scaling and makes it hard to maintain.
The post also notes that IPVS scheduler choices give little benefit here, because each node picks endpoints independently. Users on Kubernetes 1.36 or later with a kernel of 5.13 or newer are advised to switch to the nftables mode, which is feature-complete and performs about as well. Those who cannot switch can move to the improved iptables backend, and the post walks through how to change the kube-proxy configuration.
Editorial note: This is a summary of an article first published by Kubernetes Blog. The full piece is on their site.