Stay up to date with AI tools, model releases, and developer workflows that matter.
Weekly. Free. One click to leave.
Sources indicate that 7 Day Free Trial. Cancel Anytime.
Industry observers note that in an on-premises Kubernetes environment, removing NAT can make the network simpler to inspect, reduce translation overhead, and let the data center route directly to workloads. But direct routing is not automatically the right choice.
Industry observers note that the decision depends on what the organization controls and what it needs from the network. A team that owns the physical fabric, has enough address space, and is comfortable operating BGP may benefit from making pod networks directly routable. A smaller environment, a team without control of the underlay, or a cluster that benefits from network abstraction may be better served by NAT or an overlay.
In a recent development, kubernetes is also changing the ground under these decisions: kube-proxy is set to drop IPVS support starting with version 1.43, with nftables as the replacement. For on-prem teams already reviewing their service data path, it is a good moment to ask whether NAT still belongs in it.
So the useful question is not how to eliminate NAT, but when the operational benefits of doing so justify the additional routing responsibility.
NAT solves a real problem. It allows an internal network to use its own address space without requiring the surrounding infrastructure to know how to reach every endpoint inside it. This is especially useful when you donโt control the underlying network.
However, an on-premises data center is different. When you control both the Kubernetes cluster and the surrounding physical network, you can make Kubernetes networks directly routable.
Recent findings suggest that in our case, removing NAT had practical and architectural benefits. It improved network performance, reduced CPU load on the nodes, and removed the risk of conntrack tables filling up on Kubernetes hosts. A directly routed L3 model also works well for traffic balancing when a node has two uplinks.
This does not mean eliminating NAT from the data center entirely. NAT can remain at the boundary of the on-premises network, for example on hardware firewalls. Inside the data center, however, Kubernetes networks can be directly routable, allowing the physical network to reach workloads using the same routing mechanisms it already uses elsewhere.
It has been revealed that in our implementation, BGP connects the Kubernetes network to the data center fabric: BIRD runs on the nodes and advertises pod CIDRs and selected service addresses, while kube-router manages Kubernetes networking. We described this implementation in more detail in our earlier article on building a routable on-prem.
It has been revealed that kubernetes cluster with BGP and ECMP. The important point for this decision is not the particular stack, but the requirement it creates: the surrounding network must be able to accept and route Kubernetes prefixes safely. If the organization cannot or does not want to make the cluster part of its routing domain, keeping NAT or an overlay is usually the simpler choice.
According to recent reports, removing NAT alone does not make a network resilient. If every node still depends on one switch, one NIC, or one uplink, direct routing simply exposes the same physical single point of failure more clearly.
Sources indicate that the resilience benefit appears when the physical topology supports multiple independent paths. In our setup, nodes connect to independent data center switches and ECMP can use equivalent routes across those paths. If one path disappears, traffic can continue over the other.
Recent findings suggest that that distinction matters when deciding whether direct routing is worth the effort. Organizations with redundant uplinks and switches can integrate Kubernetes into an already resilient routed fabric. If the underlay itself has no path redundancy, removing NAT should not be presented as a high-availability improvement: the physical failure domain still dominates the design.
Recent findings suggest that our stack is one example of how direct routing can be implemented, not a default recommendation for clusters that remove NAT. We chose kube-router because it supports a minimal installation as a binary and systemd service and covers the networking functions we need for this L3 design. BIRD handles route advertisements to the surrounding network.
Other environments may reasonably choose Calico for conventional L3 routing or Cilium when an eBPF-based dataplane and its additional capabilities are useful. The choice of CNI and routing daemon is secondary to the architectural decision: first decide whether exposing Kubernetes prefixes to the physical network makes operational sense, then choose the implementation that fits the teamโs requirements.
In a recent development, A directly routed Kubernetes network is not automatically better than an overlay or NAT-based architecture. Each model solves a different operational problem.
Recent findings suggest that NAT is a reasonable choice when the cluster needs to remain isolated from the surrounding address strategy, when pod addresses cannot be routed through the data center, or when the infrastructure team does not control the upstream network. It also keeps Kubernetes-specific prefixes out of the physical routing domain, which can be valuable for smaller environments where direct visibility into every pod is not vital.
In a recent development, an overlay is useful when portability and abstraction matter more than exposing the physical packet path. It lets Kubernetes networking remain largely independent of the underlay and can simplify deployment across heterogeneous infrastructure. The tradeoff is another layer to inspect during troubleshooting and, depending on the implementation, encapsulation overhead.
Sources indicate that direct routing makes the most sense when the organization owns the physical network, has address space available for pod CIDRs, and has operators comfortable with BGP and route filtering. It becomes more attractive when preserving source addresses, using standard network troubleshooting tools, predictable performance, or integrating the cluster with redundant data center paths are key. At a larger scale, those operational benefits can justify the extra coordination between Kubernetes and network teams.
The broader lesson is not that every Kubernetes cluster needs BGP or that NAT should disappear everywhere. It is that Kubernetes has matured enough that we should stop automatically treating its network as a black box.
As more production and AI workloads move onto Kubernetes, network architecture becomes part of application reliability. In an on-premises data center, BGP and ECMP give us the option to use the resilience already available in the physical network rather than building another isolated networking layer on top of it.
Sometimes the best way to make Kubernetes networking more robust is not to add another abstraction, but to remove one. But only when the surrounding infrastructure and the team operating it are ready to take responsibility for the routing that abstraction previously hid.
According to recent reports, if most of those answers are yes, direct routing is worth serious consideration. If not, NAT or an overlay may be the more resilient operational choiceโnot because it is technically more advanced, but because it better matches the environment.
7 Day Free Trial. Cancel Anytime.
