kube-proxy IPVS Mode Stops Working by Default in Kubernetes 1.40
On 16 November 2026 Kubernetes SIG Network published a migration notice for kube-proxy's ipvs backend. It has been deprecated since 1.35, and the post says SIG Network now plans "to remove it in Kubernetes 1.40." Every ipvs user, it says, "should therefore be planning to migrate to either the nftables or iptables backends."
That date needs a closer reading, because the design document behind it says something slightly different, and the difference decides whether your 1.40 upgrade fails.
What happened
The post explains why ipvs is going. It was written to fix iptables' scaling problems, and it did, but the kernel IPVS API "turned out to be a bad match for the Kubernetes Services API." It never covered every Service edge case, it still leans on iptables for anything beyond load balancing, and the IPVS API has no batch updates, so big changes can be slower than with iptables. On top of that, Red Hat's plan to drop the iptables API from RHEL takes the ipvs backend with it.
The timeline is set in KEP-5495, last updated in May 2026:
| Release | What happens to ipvs mode |
|---|---|
| 1.35 | Deprecated. kube-proxy prints a warning |
| 1.37 | KubeProxyIPVS feature gate added, on by default |
| 1.40 | Gate off by default. In the KEP's words, kube-proxy "will exit with an error" unless you turn the gate back on |
| 1.43 | Gate locked off and the ipvs code removed |
| 1.46 | The gate itself removed |
So "removed in 1.40" in the blog means "stops working unless you opt back in". The code goes in 1.43. kube-proxy 1.37 says the same in its own log: "disabled by default in Kubernetes 1.40 and removed in Kubernetes 1.43."
No 1.40 date exists yet. 1.38 is planned for 16 December 2026, and on the usual three-releases-a-year rhythm 1.40 would land in the second half of 2027. That is my estimate, not a schedule. The Kubernetes release calendar row is updated with the KEP dates.
What changes for engineers
The recommended target is nftables. The post recommends nftables mode for clusters on Kubernetes 1.36 or later with kernel 5.13 or later. It has been GA since 1.33, but scaling fixes after that were not backported below 1.36. SIG Network calls it "fully supported, feature-complete", with latency slightly better than ipvs. The original nftables mode post has the design and benchmarks.
iptables is the fallback, with a reboot. If you are on an older release or kernel, the post says iptables is "not as slow as you think" after years of optimisation, including partial syncs. But it recommends rebooting each node after switching to iptables, not just restarting kube-proxy. In my own test, skipping that reboot left a deleted Service still answering (details below).
IPVS schedulers were never doing what people thought. A common reason to pick ipvs was least-connection or Maglev scheduling. The post explains that each node's kube-proxy schedules alone, without the other nodes' state, so in practice those schedulers do no better than random selection. If you set ipvs.scheduler, losing it should cost you nothing.
Managed services mostly default to iptables already. Azure AKS documents a kube-proxy configuration (still in preview) whose mode can be IPTABLES, IPVS or NFTABLES and "defaults to IPTABLES", and its page says nftables "is recommended as a replacement for IPVS." Upstream kube-proxy's default on Linux is iptables too, per the virtual IPs and proxy modes page. The clusters at risk are the ones someone deliberately set to ipvs, often years ago, for scale.
What I ran
I ran the migration on a local two-node kind cluster (Kubernetes 1.37.0) with 200 Services. I switched ipvs to nftables, then to iptables, and turned the KubeProxyIPVS gate off to see the 1.40 behaviour early. The full config, sync timings and failures are in kube-proxy ipvs to nftables, tested. In short:
| Check | Result on kind 1.37.0 |
|---|---|
ipvs mode with KubeProxyIPVS: false (the 1.40 default) |
kube-proxy crash-loops: "must be iptables, nftables or blank" |
| ipvs to nftables, kube-proxy restart only | Worked. kube-proxy removed the kube-ipvs0 interface and IPVS state |
| ipvs to iptables, kube-proxy restart only | IPVS state left behind. A deleted Service kept answering on that node until the node restarted |
| Mean rule sync on a 100-Service burst | ipvs 108 to 119 ms, nftables 24 to 31 ms |
The timings are from a laptop VM and only compare modes against each other on the same box.
What to do this week
- Find ipvs clusters. On each cluster, check the kube-proxy log for
"Using ipvs Proxier", or look for akube-ipvs0interface on a node. On kubeadm clusters,kubectl -n kube-system get cm kube-proxy -o yaml | grep "mode:"shows the setting. - Check kernel and version. nftables mode wants Kubernetes 1.36 or later and kernel 5.13 or later. Nodes below that need iptables, or a node image update first.
- Pick nftables where you can, and roll it one node pool at a time. On kubeadm, change
mode: ipvstomode: nftablesin the kube-proxy ConfigMap and restart the kube-proxy pods, as in the kubeadm reconfigure guide. Move any overridden ipvs options to their nftables equivalents. - If you pick iptables, reboot. Do not leave nodes running with old IPVS virtual servers after a restart-only switch.
- Do not plan on the gate. Turning
KubeProxyIPVSback on in 1.40 buys at most three releases, until 1.43.
FAQ
Is kube-proxy ipvs mode removed in Kubernetes 1.40?
It stops working by default in 1.40: the KubeProxyIPVS feature gate turns off, and kube-proxy in ipvs mode exits with an error unless you turn the gate back on. KEP-5495 removes the code in 1.43.
What should I replace kube-proxy ipvs mode with?
nftables mode, on Kubernetes 1.36 or later and Linux kernel 5.13 or later. If you cannot meet those, use iptables mode and reboot each node after switching.
Does switching kube-proxy modes need downtime?
It is a ConfigMap change and a kube-proxy restart. In my kind test, existing Services kept working while kube-proxy restarted. Services created during the gap only worked once the new kube-proxy was up. Roll it one node pool at a time in production.
Related: Kubernetes Networking Explained: CNI, Services and kube-proxy, How to Upgrade a kubeadm Cluster and eBPF Explained.
Spot an error or something out of date? Email me and I will fix it and log it on the corrections page.
I'm Ajay Kumar — I build and operate PandaStack, an open-source Firecracker microVM cloud for AI agents, and write about the infrastructure and tooling around it.
Need this kind of infrastructure work? See what I do or email hello@ajayk.sh.
Related
kube-proxy ipvs to nftables, Tested: What Broke on Kubernetes 1.37
I switched kube-proxy from ipvs to nftables and to iptables on a kind 1.37 cluster with 200 Services, and turned off the KubeProxyIPVS gate to preview 1.40. Real config, sync timings, a deleted Service that kept answering, and the fixes.
11 minNov 10, 2026How to Fix AWS ALB Unhealthy Targets and 502/503 Errors (2026)
Fix AWS Application Load Balancer unhealthy targets and 502, 503 and 504 errors step by step: read the health reason code, fix health check paths and success codes, security groups, ip vs instance targets, EKS readiness gates, deregistration delay, keep-alive timeouts and access logs.
13 minNov 10, 2026Kubernetes Networking Explained (2026): CNI, Services, kube-proxy in 1.37
How Kubernetes networking works in 2026: the flat pod network model, what a CNI plugin does (Cilium 1.20, Calico 3.33, Flannel 0.28), ClusterIP vs NodePort vs LoadBalancer, EndpointSlices, cluster DNS, and why kube-proxy is moving from iptables and IPVS to nftables or eBPF.
11 min