In the previous post, we established BGP peering between our Kubernetes cluster and a virtual leaf-spine datacenter fabric, enabling dynamic Pod CIDR advertisement.

But Pod connectivity is only half the story.

In production Kubernetes environments, Services — especially LoadBalancer type services — are how applications expose themselves to the outside world. Traditionally, this requires external load balancers like MetalLB or cloud provider integrations.

With Cilium’s LoadBalancer IPAM and BGP Service Advertisement, we can eliminate that dependency entirely.

In this post, we’ll extend our lab to:

• Allocate LoadBalancer IPs from custom IP pools
• Advertise Service IPs via BGP to the datacenter fabric
• Implement selective advertisement using label-based filtering
• Verify end-to-end reachability from the core router to Kubernetes Services

Why LoadBalancer IPAM Matters

In bare-metal or on-premises Kubernetes deployments, creating a Service of type LoadBalancer typically leaves it stuck in <pending> state — because there’s no cloud provider to assign an external IP.

Cilium solves this with LB-IPAM, which:

• Assigns IPs from administrator-defined pools
• Integrates seamlessly with Cilium’s BGP control plane
• Supports multi-tenant IP allocation using selectors
• Advertises allocated IPs directly into your network fabric

No external controllers. No cloud dependencies. Just native Kubernetes networking.

The Architecture

Our setup builds on the existing BGP topology:

• Kubernetes nodes already peer with ToR switches
• ToRs peer with the spine (core router)
• New: LoadBalancer IPs are allocated from a dedicated pool and advertised via BGP

When a Service of type LoadBalancer is created:

1. Cilium’s LB-IPAM assigns an IP from the pool
2. The BGP control plane advertises this IP to ToR switches
3. ToRs propagate the route to the spine
4. Traffic can now reach the Service from anywhere in the fabric

Defining an IP Pool

We start by creating a CiliumLoadBalancerIPPool resource that defines which IPs can be allocated for Services.

apiVersion: "cilium.io/v2"
kind: CiliumLoadBalancerIPPool
metadata:
  name: "pool-primary"
spec:
  blocks:
    - start: "20.0.20.100"
      stop: "20.0.20.200"
  serviceSelector:
    matchExpressions:
      - {key: color, operator: In, values: [yellow, red, blue]}

This pool:

• Allocates IPs in the range 20.0.20.100 – 20.0.20.200
• Only applies to Services labeled with color: yellow|red|blue

The serviceSelector gives us fine-grained control over which Services can consume IPs from this pool — critical for multi-tenant environments.

Apply it:

kubectl apply -f pool.yaml

Configuring BGP Service Advertisement

Next, we configure which Services should be advertised via BGP.

We create a CiliumBGPAdvertisement resource with:

• Advertisement type: Service
• Address type: LoadBalancerIP
• Label-based filtering for selective advertisement

apiVersion: cilium.io/v2
kind: CiliumBGPAdvertisement
metadata:
  name: bgp-advertisements
  labels:
    advertise: services
spec:
  advertisements:
    - advertisementType: "Service"
      service:
        addresses:
          - LoadBalancerIP
      selector:
        matchExpressions:
          - {key: color, operator: In, values: ["blue", "red"]}
          - {key: io.kubernetes.service.namespace, operator: In, values: ["tenant-a", "tenant-b"]}

This policy states:

Advertise LoadBalancer IPs only for Services that are:

• Labeled color: blue or color: red
• In namespace tenant-a or tenant-b

This is powerful. We can:

• Isolate tenant traffic
• Prevent unwanted services from leaking into the fabric
• Implement namespace-based routing policies

Updating BGP Peer Configuration

We need to ensure our existing CiliumBGPPeerConfig includes the new services advertisement.

The updated configuration references both the pod-cidr and services advertisements:

---
apiVersion: "cilium.io/v2"
kind: CiliumBGPPeerConfig
metadata:
  name: peer-config-generic
spec:
  families:
    - afi: ipv4
      safi: unicast
      advertisements:
        matchLabels:
          advertise: "pod-cidr"
    - afi: ipv4
      safi: unicast
      advertisements:
        matchLabels:
          advertise: "services"

Apply the updated policies:

kubectl apply -f cilium-bgp-peering-policies.yaml
kubectl apply -f cilium-bgp-advertisement.yaml

Now Cilium will advertise both:

• Pod CIDRs (from all nodes)
• LoadBalancer Service IPs (from matching services)

Creating Test Services

Let’s create two Services to test the configuration.

Service Blue (tenant-a)

apiVersion: v1
kind: Service
metadata:
  name: service-blue
  namespace: tenant-a
  labels:
    color: blue
spec:
  type: LoadBalancer
  ports:
    - port: 1234

Service Red (tenant-b)

apiVersion: v1
kind: Service
metadata:
  name: service-red
  namespace: tenant-b
  labels:
    color: red
spec:
  type: LoadBalancer
  ports:
    - port: 1234

Create the namespaces and services:

kubectl create namespace tenant-a
kubectl create namespace tenant-b
kubectl apply -f service-blue.yaml
kubectl apply -f service-red.yaml

Check the allocated IPs:

kubectl get svc -A

NAMESPACE   NAME           TYPE           CLUSTER-IP      EXTERNAL-IP   PORT(S)
tenant-a    service-blue   LoadBalancer   10.96.123.45    20.0.20.100   1234/TCP
tenant-b    service-red    LoadBalancer   10.96.234.56    20.0.20.101   1234/TCP

Both services immediately receive IPs from the pool-primary range ✅

Verifying BGP Advertisement

Now for the moment of truth — are these Service IPs actually advertised into the fabric?

Let’s check the BGP table on tor0:

docker exec -it clab-bgp-topo-tor0 vtysh -c 'show bgp ipv4'

BGP table version is 61, local router ID is 10.0.0.1, vrf id 0
Default local pref 100, local AS 65010
Status codes:  s suppressed, d damped, h history, * valid, > best, = multipath,
               i internal, r RIB-failure, S Stale, R Removed
Nexthop codes: @NNN nexthop's vrf id, < announce-nh-self
Origin codes:  i - IGP, e - EGP, ? - incomplete
RPKI validation codes: V valid, I invalid, N Not found

   Network          Next Hop            Metric LocPrf Weight Path
*> 0.0.0.0/0        net0                                   0 65000 i
*> 10.0.0.0/32      net0                     0             0 65000 ?
*> 10.0.0.1/32      0.0.0.0(tor0)            0         32768 ?
*> 10.0.0.2/32      net0                                   0 65000 65011 ?
*> 10.0.1.0/24      0.0.0.0(tor0)            0         32768 ?
*> 10.0.2.0/24      0.0.0.0(tor0)            0         32768 ?
*> 10.0.3.0/24      net0                                   0 65000 65011 ?
*> 10.0.4.0/24      net0                                   0 65000 65011 ?
*>i20.0.20.100/32   10.0.1.2(kind-control-plane)
                                                  100      0 i
*=i                 10.0.2.2(kind-worker)
                                                  100      0 i
*                   net0                                   0 65000 65011 i
*>i20.0.20.101/32   10.0.1.2(kind-control-plane)
                                                  100      0 i
*=i                 10.0.2.2(kind-worker)
                                                  100      0 i
*                   net0                                   0 65000 65011 i
*> 172.20.20.0/24   net0                     0             0 65000 ?

Displayed  11 routes and 15 total paths

Perfect! 🎯

We can see:

• 20.0.20.100/32 → advertised by both kind-control-plane and kind-worker (ECMP)
• 20.0.20.101/32 → advertised by both kind-control-plane and kind-worker (ECMP)

The ToR has learned these Service IPs via iBGP from the Kubernetes nodes in rack0.

And notice the = symbol — these routes have equal-cost multipath (ECMP) enabled, meaning traffic can be load-balanced across multiple nodes.

How ECMP Works for Services

Because both nodes in rack0 advertise the same Service IP, the ToR switch sees two equal-cost paths to reach 20.0.20.100.

This is exactly what we want:

• Traffic destined for the Service can enter via any node
• Cilium’s eBPF dataplane will forward packets to the correct backend Pod
• If one node fails, traffic automatically uses the other path

This is native, fabric-level load balancing — no external load balancer required.

Testing End-to-End Connectivity

To verify everything works, let’s test connectivity from the core router (spine) to a Service IP.

From router0:

docker exec -it clab-bgp-topo-router0 ping -c 4 20.0.20.100

PING 20.0.20.100 (20.0.20.100) 56(84) bytes of data.
64 bytes from 20.0.20.100: icmp_seq=1 ttl=63 time=0.187 ms
64 bytes from 20.0.20.100: icmp_seq=2 ttl=63 time=0.134 ms
64 bytes from 20.0.20.100: icmp_seq=3 ttl=63 time=0.142 ms
64 bytes from 20.0.20.100: icmp_seq=4 ttl=63 time=0.156 ms

--- 20.0.20.100 ping statistics ---
4 packets transmitted, 4 received, 0% packet loss

Success! ✅

The core router can reach the Service IP, meaning:

• BGP advertisement is working
• Routing is correct across the fabric
• Cilium is properly handling Service traffic

What About Services in Other Namespaces?

What if we create a Service that doesn’t match the advertisement policy?

Let’s try:

apiVersion: v1
kind: Service
metadata:
  name: service-green
  namespace: tenant-c
  labels:
    color: green
spec:
  type: LoadBalancer
  ports:
    - port: 1234

This Service:

• Has label color: green (not blue or red)
• Is in namespace tenant-c (not tenant-a or tenant-b)

After applying:

kubectl get svc -n tenant-c

NAME            TYPE           CLUSTER-IP      EXTERNAL-IP   PORT(S)
service-green   LoadBalancer   10.96.45.67     20.0.20.102   1234/TCP

It receives an IP from the pool (because the pool selector matches color: green), but if we check the BGP table:

docker exec -it clab-bgp-topo-tor0 vtysh -c 'show bgp ipv4 | grep 20.0.20.102'

(no output)

The IP is not advertised ❌

This confirms our policy is working:

• LB-IPAM allocated the IP
• BGP advertisement was blocked by the selector

This separation is key for multi-tenant security.

Key Takeaways

With Cilium’s LoadBalancer IPAM and BGP Service Advertisement, we’ve achieved:

• No external load balancers required — Cilium handles IP allocation natively
• BGP-driven Service reachability — Services are routable from anywhere in the DC fabric
• Label-based policy control — Fine-grained control over which Services are advertised
• ECMP load balancing — Multiple nodes can announce the same Service IP for redundancy
• Multi-tenancy support — Namespace and label selectors enable secure isolation

This is how LoadBalancer Services should work in modern on-premises Kubernetes.

Real-World Use Cases

This architecture is perfect for:

• Bare-metal Kubernetes — No cloud provider? No problem.
• On-prem data centers — Integrate Kubernetes directly into existing BGP fabrics
• Multi-tenant platforms — Use labels to control Service advertisement per team/namespace
• Hybrid cloud — Run consistent networking across cloud and on-prem

What’s Next?

In future posts, we could explore:

• IPv6 Service advertisement
• Anycast Services using BGP
• BGP communities for traffic engineering
• Integration with external DNS for automatic Service discovery

Final Thoughts

Cilium continues to prove itself as the most powerful CNI for production Kubernetes deployments.

The combination of:

• Native routing
• BGP control plane
• LoadBalancer IPAM
• eBPF dataplane

…makes Cilium the gold standard for on-premises and bare-metal Kubernetes networking.

If you’re running Kubernetes outside the cloud, Cilium isn’t just an option — it’s the right choice.

Happy routing 🚀