The Kubernetes project announced on November 16, 2026 that the ipvs mode of kube-proxy is scheduled for removal in Kubernetes 1.40. The notice came from SIG Network, the group that maintains cluster networking, in a post by Dan Winship of Red Hat.
The ipvs mode has been deprecated since Kubernetes 1.35 and still runs today with a warning. Clusters that use it now have a fixed deadline to switch to the nftables or iptables mode, and the change is mostly a configuration edit.
What Was Announced
Kube-proxy is the component on every node that turns Kubernetes Services into packet forwarding rules. It has three Linux backends: iptables, ipvs and nftables. The nftables backend arrived in Kubernetes 1.29 and reached general availability in 1.33, with the stated aim of replacing the other two over time.
SIG Network gives three reasons for dropping ipvs first. The code is complex, it never covered every edge case of Service behavior, and its design differs so much from the other backends that it costs more to maintain.
Why IPVS Was Added and Why It Fell Short
The backend was created in 2017 to fix two scaling limits of the iptables mode. New connections were checked against one rule per Service in sequence, so latency grew with the number of Services. Worse, kube-proxy rewrote the whole ruleset on every Service or endpoint change, which could pin a CPU core in very large clusters.
IPVS fixed both. The kernel subsystem looks up Services in hash tables, and updates touch only what changed. Over time, though, other problems appeared:
- The kernel IPVS interface has no batch update call, so a large burst of changes can be slower than with iptables.
- IPVS only does load balancing, so the backend still leans on iptables and ipset for other Service features and inherits some of the same limits.
- The first authors left the project, and it took time to find new maintainers.
- Red Hat plans to remove the iptables interface from Red Hat Enterprise Linux, which would make the ipvs mode unusable there as well.
The Scheduler Argument
Some teams chose ipvs for its selectable schedulers, hoping for smarter traffic distribution than random choice. The post explains that this benefit is mostly an illusion in kube-proxy.
Each node runs its own kube-proxy and none of them share scheduler state. A node picks endpoints with no knowledge of what clients on other nodes are doing. Even Maglev hashing, which is built for distributed load balancers, needs a two-tier design that kube-proxy does not have.
How to Migrate
The recommended target is nftables. SIG Network advises it for clusters on Kubernetes 1.36 or later whose nodes run Linux kernel 5.13 or newer. The backend has been usable since 1.33, but some performance and scaling fixes were not backported to releases before 1.36.
For kubeadm clusters, the switch means changing the mode value from ipvs to nftables in the kube-proxy ConfigMap and restarting the kube-proxy pods. Any ipvs-specific options you set must be replaced with their nftables equivalents. The project's own latency chart shows nftables performing marginally better than ipvs.
Clusters on older Kubernetes versions or older distributions can fall back to iptables instead. That backend has improved a lot since 2017:
- Redundant rules were removed from the generated ruleset.
- The generator allocates less memory and avoids repeated computation.
- Kube-proxy no longer reads the full ruleset before each sync, which was very slow at scale.
- Only rules for changed Services and endpoints are rewritten where the iptables interface allows it.
When moving from ipvs to iptables, the project recommends rebooting each node after the configuration change, not only restarting kube-proxy.
What It Means for Developers
Application developers should see no difference. Service addresses, DNS names and ports behave the same in every backend, so no manifest changes are needed for normal workloads.
Platform teams have real work to plan. Start by checking which mode each cluster runs, since the setting may have been chosen years ago and forgotten. Users of managed Kubernetes should read their provider's notes, because the provider may control this setting.
Test the move in a staging cluster first. Look for anything that depends on ipvs details, such as a chosen scheduler or monitoring that reads IPVS counters. Teams that run databases inside Kubernetes, a pattern also behind the recently launched CoreWeave managed PostgreSQL service, should confirm that long-lived connections survive the node-by-node rollout.
Do not leave it for the 1.40 upgrade itself. Changing the proxy mode and the Kubernetes version in the same maintenance window makes any networking fault much harder to trace.
Conclusion
The ipvs backend solved a real scaling problem in its day, but nftables now does the same job with a simpler design and full feature coverage. With removal set for Kubernetes 1.40, clusters still on ipvs should schedule the move to nftables now. The full timeline is in KEP-5495.
Sources
See my articles more often when you search Google.
