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.


  • Expose a simple application through various types of Services
  • Understand how each Service type handles source IP NAT
  • Understand the tradeoffs involved in preserving source IP

Before you begin


This document makes use of the following terms:

network address translation
Source NAT
replacing the source IP on a packet; in this page, that usually means replacing with the IP address of a node.
Destination NAT
replacing the destination IP on a packet; in this page, that usually means replacing with the IP address of a PodThe smallest and simplest Kubernetes object. A Pod represents a set of running containers on your cluster.
a virtual IP address, such as the one assigned to every ServiceA way to expose an application running on a set of Pods as a network service. in Kubernetes
a network daemon that orchestrates Service VIP management on every node


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:

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 create deployment source-ip-app --image=k8s.gcr.io/echoserver:1.4

The output is:

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, (the default). You can query the kube-proxy mode by fetching http://localhost:10249/proxyMode on the node where kube-proxy is running.

kubectl get nodes

The output is similar to this:

NAME                           STATUS     ROLES    AGE     VERSION
kubernetes-node-6jst   Ready      <none>   2h      v1.13.0
kubernetes-node-cx31   Ready      <none>   2h      v1.13.0
kubernetes-node-jj1t   Ready      <none>   2h      v1.13.0

Get the proxy mode on one of the nodes (kube-proxy listens on port 10249):

# Run this in a shell on the node you want to query.
curl http://localhost:10249/proxyMode

The output is:


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

The output is:

service/clusterip exposed
kubectl get svc clusterip

The output is similar to:

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

The output is similar to this:

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.

You can then run a command inside that Pod:

# Run this inside the terminal from "kubectl run"
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

…then use wget to query the local webserver

# Replace with the Pod's IPv4 address
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

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

The output is:

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 cloud provider, 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

The output is similar to:


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

  • Client sends packet to node2:nodePort
  • node2 replaces the source IP address (SNAT) in the packet with its own IP address
  • node2 replaces the destination IP on the packet with the pod IP
  • packet is routed to node 1, and then to the endpoint
  • the pod’s reply is routed back to node2
  • the pod’s reply is sent back to the client


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

To avoid this, Kubernetes has a feature to preserve the client source IP. If you set service.spec.externalTrafficPolicy to the value Local, kube-proxy only proxies proxy requests to local endpoints, and does not forward traffic to other nodes. This approach preserves 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"}}'

The output is:

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

The output is similar to:


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:

  • client sends packet to node2:nodePort, which doesn’t have any endpoints
  • packet is dropped
  • client sends packet to node1:nodePort, which does have endpoints
  • node1 routes packet to endpoint with the correct source IP


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

Source IP for Services with Type=LoadBalancer

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 load-balanced 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 load balancer:

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

The output is:

service/loadbalancer exposed

Print out the IP addresses of the Service:

kubectl get svc loadbalancer

The output is similar to this:

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

Next, send a request to this Service’s external-ip:


The output is similar to this:


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

The output is similar to this:

  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

The output is similar to this:

NAME                            READY     STATUS    RESTARTS   AGE       IP             NODE
source-ip-app-826191075-qehz4   1/1       Running   0          20h   kubernetes-node-6jst

Use curl to fetch the /healthz endpoint on various nodes:

# Run this locally on a node you choose
curl localhost:32122/healthz
1 Service Endpoints found

On a different node you might get a different result:

# Run this locally on a node you choose
curl localhost:32122/healthz
No Service Endpoints Found

A controller running on the control planeThe container orchestration layer that exposes the API and interfaces to define, deploy, and manage the lifecycle of containers. is responsible for allocating the cloud load balancer. The same controller 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 use curl to query the IPv4 address of the load balancer:


The output is similar to this:


Cross-platform support

Only some cloud providers offer support for source IP preservation through Services with Type=LoadBalancer. The cloud provider 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.

Load balancers 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 Forwarded or X-FORWARDED-FOR headers, or the proxy protocol. Load balancers in the second category can leverage the feature described above by 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