Kubernetes plans to remove kube-proxy's IPVS backend in 1.40 and asks users to move to nftables
Kubernetes SIG Network now plans to remove the IPVS backend of kube-proxy in Kubernetes 1.40, Dan Winship of Red Hat wrote on the Kubernetes blog on 16 November. The backend was deprecated in 1.35 and still works, with a warning; all IPVS users should plan to move to the nftables or iptables backend.
The IPVS backend was written to fix two scaling problems of the iptables backend: connection setup that had to walk one rule per service, and kube-proxy rewriting the whole ruleset on every change, which in very large clusters kept a CPU busy all the time. IPVS solved both with hash tables and incremental updates. But the post says the kernel IPVS API turned out to be a bad match for the Kubernetes Services API. The backend never implemented every edge case, its original authors moved on, and with no batch update operation it could be slower than iptables when many things change at once. It also falls back to iptables for features beyond basic load balancing, so it shares the iptables backend's long-term problem: Red Hat plans to remove the iptables API from Red Hat Enterprise Linux.
The IPVS schedulers some users wanted do not help either, the post says, because kube-proxy on each node picks endpoints without knowing what other nodes are doing; even Maglev hashing does not beat random selection in that setting.
For clusters on Kubernetes 1.36 or later with Linux kernel 5.13 or later, the project recommends the nftables mode, GA since 1.33 but with performance fixes not backported before 1.36. It is feature-complete and, in the project's measurements, very slightly faster than IPVS. In a kubeadm cluster the switch is a change from mode: ipvs to mode: nftables in the kube-proxy ConfigMap and a restart of the kube-proxy pods; IPVS-specific options need their nftables counterparts. Where nftables is not possible, the iptables backend has been optimised a great deal since 2017.

Why it matters
The deprecation has a date now: 1.40. Clusters still on IPVS have a supported route and a version floor - 1.36 and kernel 5.13 for nftables - and the post explains why staying on IPVS is not an option once the iptables API it leans on leaves RHEL.