Edit This Page

Using Source IP

Applications running in a Kubernetes cluster find and communicate with each other, and the outside world, through the Service abstraction. This document explains what happens to the source IP of packets sent to different types of Services, and how you can toggle this behavior according to your needs.


Before you begin

You need to have a Kubernetes cluster, and the kubectl command-line tool must be configured to communicate with your cluster. If you do not already have a cluster, you can create one by using Minikube, or you can use one of these Kubernetes playgrounds:

To check the version, enter kubectl version.


This document makes use of the following terms:


You must have a working Kubernetes 1.5 cluster to run the examples in this document. The examples use a small nginx webserver that echoes back the source IP of requests it receives through an HTTP header. You can create it as follows:

$ kubectl run source-ip-app --image=k8s.gcr.io/echoserver:1.4
deployment.apps/source-ip-app created

Source IP for Services with Type=ClusterIP

Packets sent to ClusterIP from within the cluster are never source NAT’d if you’re running kube-proxy in iptables mode, which is the default since Kubernetes 1.2. Kube-proxy exposes its mode through a proxyMode endpoint:

$ kubectl get nodes
NAME                           STATUS     ROLES    AGE     VERSION
kubernetes-minion-group-6jst   Ready      <none>   2h      v1.13.0
kubernetes-minion-group-cx31   Ready      <none>   2h      v1.13.0
kubernetes-minion-group-jj1t   Ready      <none>   2h      v1.13.0

kubernetes-minion-group-6jst $ curl localhost:10249/proxyMode

You can test source IP preservation by creating a Service over the source IP app:

$ kubectl expose deployment source-ip-app --name=clusterip --port=80 --target-port=8080
service/clusterip exposed

$ kubectl get svc clusterip
clusterip    ClusterIP   <none>        80/TCP    51s

And hitting the ClusterIP from a pod in the same cluster:

$ kubectl run busybox -it --image=busybox --restart=Never --rm
Waiting for pod default/busybox to be running, status is Pending, pod ready: false
If you don't see a command prompt, try pressing enter.

# ip addr
1: lo: <LOOPBACK,UP,LOWER_UP> mtu 65536 qdisc noqueue
    link/loopback 00:00:00:00:00:00 brd 00:00:00:00:00:00
    inet scope host lo
       valid_lft forever preferred_lft forever
    inet6 ::1/128 scope host
       valid_lft forever preferred_lft forever
3: eth0: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1460 qdisc noqueue
    link/ether 0a:58:0a:f4:03:08 brd ff:ff:ff:ff:ff:ff
    inet scope global eth0
       valid_lft forever preferred_lft forever
    inet6 fe80::188a:84ff:feb0:26a5/64 scope link
       valid_lft forever preferred_lft forever

# wget -qO -

The client_address is always the client pod’s IP address, whether the client pod and server pod are in the same node or in different nodes.

Source IP for Services with Type=NodePort

As of Kubernetes 1.5, packets sent to Services with Type=NodePort are source NAT’d by default. You can test this by creating a NodePort Service:

$ kubectl expose deployment source-ip-app --name=nodeport --port=80 --target-port=8080 --type=NodePort
service/nodeport exposed

$ NODEPORT=$(kubectl get -o jsonpath="{.spec.ports[0].nodePort}" services nodeport)
$ NODES=$(kubectl get nodes -o jsonpath='{ $.items[*].status.addresses[?(@.type=="ExternalIP")].address }')

If you’re running on a cloudprovider, you may need to open up a firewall-rule for the nodes:nodeport reported above. Now you can try reaching the Service from outside the cluster through the node port allocated above.

$ for node in $NODES; do curl -s $node:$NODEPORT | grep -i client_address; done

Note that these are not the correct client IPs, they’re cluster internal IPs. This is what happens:


             \ ^
              \ \
               v \
   node 1 <--- node 2
    | ^   SNAT
    | |   --->
    v |

To avoid this, Kubernetes has a feature to preserve the client source IP (check here for feature availability). Setting service.spec.externalTrafficPolicy to the value Local will only proxy requests to local endpoints, never forwarding traffic to other nodes and thereby preserving the original source IP address. If there are no local endpoints, packets sent to the node are dropped, so you can rely on the correct source-ip in any packet processing rules you might apply a packet that make it through to the endpoint.

Set the service.spec.externalTrafficPolicy field as follows:

$ kubectl patch svc nodeport -p '{"spec":{"externalTrafficPolicy":"Local"}}'
service/nodeport patched

Now, re-run the test:

$ for node in $NODES; do curl --connect-timeout 1 -s $node:$NODEPORT | grep -i client_address; done

Note that you only got one reply, with the right client IP, from the one node on which the endpoint pod is running.

This is what happens:


       ^ /   \
      / /     \
     / v       X
   node 1     node 2
    ^ |
    | |
    | v

Source IP for Services with Type=LoadBalancer

As of Kubernetes 1.5, packets sent to Services with Type=LoadBalancer are source NAT’d by default, because all schedulable Kubernetes nodes in the Ready state are eligible for loadbalanced traffic. So if packets arrive at a node without an endpoint, the system proxies it to a node with an endpoint, replacing the source IP on the packet with the IP of the node (as described in the previous section).

You can test this by exposing the source-ip-app through a loadbalancer

$ kubectl expose deployment source-ip-app --name=loadbalancer --port=80 --target-port=8080 --type=LoadBalancer
service/loadbalancer exposed

$ kubectl get svc loadbalancer
NAME           TYPE           CLUSTER-IP    EXTERNAL-IP       PORT(S)   AGE
loadbalancer   LoadBalancer   80/TCP    5m

$ curl

However, if you’re running on Google Kubernetes Engine/GCE, setting the same service.spec.externalTrafficPolicy field to Local forces nodes without Service endpoints to remove themselves from the list of nodes eligible for loadbalanced traffic by deliberately failing health checks.


                      lb VIP
                     / ^
                    v /
health check --->   node 1   node 2 <--- health check
        200  <---   ^ |             ---> 500
                    | V

You can test this by setting the annotation:

$ kubectl patch svc loadbalancer -p '{"spec":{"externalTrafficPolicy":"Local"}}'

You should immediately see the service.spec.healthCheckNodePort field allocated by Kubernetes:

$ kubectl get svc loadbalancer -o yaml | grep -i healthCheckNodePort
  healthCheckNodePort: 32122

The service.spec.healthCheckNodePort field points to a port on every node serving the health check at /healthz. You can test this:

$ kubectl get pod -o wide -l run=source-ip-app
NAME                            READY     STATUS    RESTARTS   AGE       IP             NODE
source-ip-app-826191075-qehz4   1/1       Running   0          20h   kubernetes-minion-group-6jst

kubernetes-minion-group-6jst $ curl localhost:32122/healthz
1 Service Endpoints found

kubernetes-minion-group-jj1t $ curl localhost:32122/healthz
No Service Endpoints Found

A service controller running on the master is responsible for allocating the cloud loadbalancer, and when it does so, it also allocates HTTP health checks pointing to this port/path on each node. Wait about 10 seconds for the 2 nodes without endpoints to fail health checks, then curl the lb ip:

$ curl

Cross platform support

As of Kubernetes 1.5, support for source IP preservation through Services with Type=LoadBalancer is only implemented in a subset of cloudproviders (GCP and Azure). The cloudprovider you’re running on might fulfill the request for a loadbalancer in a few different ways:

  1. With a proxy that terminates the client connection and opens a new connection to your nodes/endpoints. In such cases the source IP will always be that of the cloud LB, not that of the client.

  2. With a packet forwarder, such that requests from the client sent to the loadbalancer VIP end up at the node with the source IP of the client, not an intermediate proxy.

Loadbalancers in the first category must use an agreed upon protocol between the loadbalancer and backend to communicate the true client IP such as the HTTP X-FORWARDED-FOR header, or the proxy protocol. Loadbalancers in the second category can leverage the feature described above by simply creating an HTTP health check pointing at the port stored in the service.spec.healthCheckNodePort field on the Service.

Cleaning up

Delete the Services:

$ kubectl delete svc -l run=source-ip-app

Delete the Deployment, ReplicaSet and Pod:

$ kubectl delete deployment source-ip-app

What's next