GKE networking overview (original) (raw)

This page provides a guide to the main aspects of Google Kubernetes Engine (GKE) networking. This information is useful to those who are just getting started with Kubernetes, as well as experienced cluster operators or application developers who need more knowledge about Kubernetes networking in order to better design applications or configure Kubernetes workloads.

This page and the rest of this documentation set are for Cloud architects and Networking specialists who design and architect the network for their organization. If you are new to Kubernetes or GKE, or need to learn more about a more general GKE topic, see the core GKE documentation, which covers core features available to all GKE users. For an overview of all GKE documentation sets, see Explore GKE documentation. To learn more about common roles and example tasks that we reference in Google Cloud content, see Common GKE Enterprise user roles and tasks.

Kubernetes lets you declaratively define how your applications are deployed, how applications communicate with each other and with the Kubernetes control plane, and how clients can reach your applications. This page also provides information about how GKE configures Google Cloud services, where it is relevant to networking.

Why Kubernetes networking is different

When you use Kubernetes to orchestrate your applications, it's important to change how you think about the network design of your applications and their hosts. With Kubernetes, you think about how Pods, Services, and external clients communicate, rather than thinking about how your hosts or virtual machines (VMs) are connected.

Kubernetes' advanced software-defined networking (SDN) enables packet routing and forwarding for Pods, Services, and nodes across different zones in the same regional cluster. Kubernetes and Google Cloud also dynamically configure IP filtering rules, routing tables, and firewall rules on each node, depending on the declarative model of your Kubernetes deployments and your cluster configuration on Google Cloud.

Prerequisites

Before reading this page, ideally ensure that you're familiar with Linux network management concepts and utilities such as iptables rules and routing.

In addition, ensure that you're familiar with basic terminology for the following topics:

The Kubernetes networking model relies heavily on IP addresses. Services, Pods, containers, and nodes communicate using IP addresses and ports. Kubernetes provides different types of load balancing to direct traffic to the correct Pods. All of these mechanisms are described in more detail later. Keep the following terms related to IP addresses in mind as you read:

Cluster connectivity requirements

All clusters require connectivity to *.googleapis.com,*.gcr.io, and the control plane IP address. This requirement is satisfied by theimplied allow egress rules and theautomatically created firewall rulesthat GKE creates.

Networking inside the cluster

This section discusses networking within a Kubernetes cluster, as it relates toIP allocation, Pods, Services,DNS, and the control plane.

IP address allocation

Kubernetes uses various IP ranges to assign IP addresses to nodes, Pods, and Services.

GKE networking model doesn't allow IP addresses to be reused across the network. When you migrate to GKE, you must plan your IP address allocation to Reduce internal IP address usage in GKE.

Maximum transmission unit (MTU)

The MTU selected for a Pod interface is dependent on the Container Network Interface (CNI) used by the cluster Nodes and the underlying VPC MTU setting. For more information, seePods.

The Pod interface MTU value is either 1460 or inherited from the primary interface of the Node.

CNI MTU GKE Standard
kubenet 1460 Default
kubenet (GKE version 1.26.1 and later) Inherited Default
Calico 1460 Enabled by using --enable-network-policy. For details, see Control communication between Pods and Services using network policies.
netd Inherited Enabled by using any of the following: Intranode visibility Workload Identity Federation for GKE IPv4/IPv6 dual-stack networking
GKE Dataplane V2 Inherited Enabled by using --enable-dataplane-v2. For details, see Using GKE Dataplane V2.

For more information, seeVPC-native clusters.

Supported network plugins

Pods

In Kubernetes, aPod is the most basic deployable unit within a Kubernetes cluster. A Pod runs one or more containers. Zero or more Pods run on a node. Each node in the cluster is part of anode pool.

In GKE, these nodes are virtual machines, each running as an instance in Compute Engine.

Pods can also attach to external storage volumes and other custom resources. This diagram shows a single node running two Pods, each attached to two volumes.

image

When Kubernetes schedules a Pod to run on a node, it creates anetwork namespacefor the Pod in the node's Linux kernel. This network namespace connects the node's physical network interface, such as eth0, with the Pod using a virtual network interface, so that packets can flow to and from the Pod. The associated virtual network interface in the node's root network namespace connects to a Linux bridge that allows communication among Pods on the same node. A Pod can also send packets outside of the node using the same virtual interface.

Kubernetes assigns an IP address (the Pod IP address) to the virtual network interface in the Pod's network namespace from a range of addresses reserved for Pods on the node. This address range is a subset of the IP address range assigned to the cluster for Pods, which you can configure when you create a cluster.

A container running in a Pod uses the Pod's network namespace. From the container's point of view, the Pod appears to be a physical machine with one network interface. All containers in the Pod see this same network interface. Each container's localhost is connected, through the Pod, to the node's physical network interface, such as eth0.

Note that this connectivity differs drastically depending on whether you use GKE's Container Network Interface (CNI) or choose to use Calico's implementation by enablingNetwork policywhen you create the cluster.

arp -n

Running the following command in the toolbox container shows the root namespace end of each veth pair attached to cbr0: 

brctl show cbr0  
arp -n

Running the following command in the toolbox container shows that there is not a Linux bridge device named cbr0: 

brctl show

The iptables rules that facilitate forwarding within the cluster differ from one scenario to the other. It is important to have this distinction in mind during detailed troubleshooting of connectivity issues.

By default, each Pod has unfiltered access to all the other Pods running on all nodes of the cluster, but you canlimit access among Pods. Kubernetes regularly tears down and recreates Pods. This happens when a node pool is upgraded, when changing the Pod's declarative configuration or changing a container's image, or when a node becomes unavailable. Therefore, a Pod's IP address is an implementation detail, and you shouldn't rely on them. Kubernetes provides stable IP addresses usingServices.

Services

In Kubernetes, you can assign arbitrary key-value pairs calledlabels to any Kubernetes resource. Kubernetes uses labels to group multiple related Pods into a logical unit called a Service. A Service has a stable IP address and ports, and provides load balancing among the set of Pods whose labels match all the labels you define in thelabel selectorwhen you create the Service.

The following diagram shows two separate Services, each of which is comprised of multiple Pods. Each of the Pods in the diagram has the label app=demo, but their other labels differ. Service "frontend" matches all Pods with bothapp=demo and component=frontend, while Service "users" matches all Pods withapp=demo and component=users. The Client Pod does not match either Service selector exactly, so it is not a part of either Service. However, the Client Pod can communicate with either of the Services because it runs in the same cluster.

image

Kubernetes assigns a stable, reliable IP address to each newly-created Service (the ClusterIP) from the cluster's pool of available Service IP addresses. Kubernetes also assigns a hostname to the ClusterIP, by adding aDNS entry. The ClusterIP and hostname are unique within the cluster and don't change throughout the lifecycle of the Service. Kubernetes only releases the ClusterIP and hostname if the Service is deleted from the cluster's configuration. You can reach a healthy Pod running your application using either the ClusterIP or the hostname of the Service.

At first glance, a Service may seem to be a single point of failure for your applications. However, Kubernetes spreads traffic as evenly as possible across the full set of Pods, running on many nodes, so a cluster can withstand an outage affecting one or more (but not all) nodes.

Kube-Proxy

Kubernetes manages connectivity among Pods and Services using the kube-proxycomponent, which traditionally runs as a static Pod on each node.

kube-proxy, which is not an inline proxy, but an egress-based load-balancing controller, watches the Kubernetes API server and continually maps the ClusterIP to healthy Pods by adding and removing destination NAT (DNAT) rules to the node's iptables subsystem. When a container running in a Pod sends traffic to a Service's ClusterIP, the node selects a Pod at random and routes the traffic to that Pod.

When you configure a Service, you can optionally remap its listening port by defining values for port and targetPort.

kube-proxy manages this port remapping by adding and removing iptables rules on the node.

This diagram illustrates the flow of traffic from a client Pod to a server Pod on a different node. The client connects to the Service at 172.16.12.100:80. The Kubernetes API server maintains a list of Pods running the application. Thekube-proxy process on each node uses this list to create an iptables rule to direct traffic to an appropriate Pod (such as 10.255.255.202:8080). The client Pod does not need to be aware of the topology of the cluster or any details about individual Pods or containers within them.

image

How kube-proxy is deployed depends on the GKE version of the cluster:

DNS

GKE provides the following managed cluster DNS options to resolve service names and external names:

GKE also providesNodeLocal DNSCacheas an optional add-on with kube-dns or Cloud DNS to improve cluster DNS performance.

To learn more about how GKE provides DNS, seeService discovery and DNS.

Control plane

In Kubernetes, the control planemanages the control plane processes, including the Kubernetes API server. How you access the control plane depends on how you've configured the control plane network isolation.

Networking outside the cluster

This section explains how traffic from outside the cluster reaches applications running within a Kubernetes cluster. This information is important when designing your cluster's applications and workloads.

You've already read about how Kubernetes uses Services to provide stable IP addresses for applications running within Pods. By default, Pods don't expose an external IP address, because kube-proxy manages all traffic on each node. Pods and their containers can communicate freely, but connections outside the cluster cannot access the Service. For example, in the previous illustration, clients outside the cluster cannot access the frontend Service using its ClusterIP.

GKE provides three different types of load balancers to control access and to spread incoming traffic across your cluster as evenly as possible. You can configure one Service to use multiple types of load balancers simultaneously.

When traffic reaches a Kubernetes node, it is handled the same way, regardless of the type of load balancer. The load balancer is not aware of which nodes in the cluster are running Pods for its Service. Instead, it balances traffic across all nodes in the cluster, even those not running a relevant Pod. On a regional cluster, the load is spread across all nodes in all zones for the cluster's region. When traffic is routed to a node, the node routes the traffic to a Pod, which may be running on the same node or a different node. The node forwards the traffic to a randomly chosen Pod by using the iptables rules that kube-proxy manages on the node.

In the following diagram, the external passthrough Network Load Balancer directs traffic to the middle node, and the traffic is redirected to a Pod on the first node.

image

When a load balancer sends traffic to a node, the traffic might get forwarded to a Pod on a different node. This requires extra network hops. If you want to avoid the extra hops, you can specify that traffic must go to a Pod that is on the same node that initially receives the traffic.

To specify that traffic must go to a Pod on the same node, setexternalTrafficPolicy to Local in your Service manifest:

apiVersion: v1
kind: Service
metadata:
  name: my-lb-service
spec:
  type: LoadBalancer
  externalTrafficPolicy: Local
  selector:
    app: demo
    component: users
  ports:
  - protocol: TCP
    port: 80
    targetPort: 8080

When you set externalTrafficPolicy to Local, the load balancer sends traffic only to nodes that have a healthy Pod that belongs to the Service. The load balancer uses a health check to determine which nodes have the appropriate Pods.

External load balancer

If your Service needs to be reachable from outside the cluster and outside your VPC network, you can configure your Service as aLoadBalancer, by setting the Service's type field to Loadbalancer when defining the Service. GKE then provisions anexternal passthrough Network Load Balancer in front of the Service. The external passthrough Network Load Balancer is aware of all nodes in your cluster and configures your VPC network's firewall rules to allow connections to the Service from outside the VPC network, using the Service's external IP address. You can assign a static external IP address to the Service.

For more information, seeConfiguring domain names with static IP addresses.

To learn more about firewall rules, seeAutomatically created firewall rules.

Technical details

When using the external load balancer, arriving traffic is initially routed to a node using a forwarding rule associated with the Google Cloud network. After the traffic reaches the node, the node uses its iptables NAT table to choose a Pod. kube-proxy manages the iptables rules on the node.

Internal load balancer

For traffic that needs to reach your cluster from within the same VPC network, you can configure your Service to provision aninternal passthrough Network Load Balancer. The internal passthrough Network Load Balancer chooses an IP address from your cluster's VPC subnet instead of an external IP address. Applications or services within the VPC network can use this IP address to communicate with Services inside the cluster.

Technical details

Internal load balancing is provided by Google Cloud. When the traffic reaches a given node, that node uses its iptables NAT table to choose a Pod, even if the Pod is on a different node.kube-proxy manages the iptables rules on the node.

For more information about internal load balancers, seeUsing an internal passthrough Network Load Balancer.

Application Load Balancer

Many applications, such as RESTful web service APIs, communicate using HTTP(S). You can allow clients external to your VPC network to access this type of application using a KubernetesIngress.

An Ingress lets you map hostnames and URL paths to Services within the cluster. When using an Application Load Balancer, you must configure the Service to use aNodePort, as well as a ClusterIP. When traffic accesses the Service on a node's IP at the NodePort, GKE routes traffic to a healthy Pod for the Service. You can specify a NodePort or allow GKE to assign a random unused port.

When you create the Ingress resource, GKE provisions anexternal Application Load Balancerin the Google Cloud project. The load balancer sends a request to a node's IP address at the NodePort. After the request reaches the node, the node uses itsiptables NAT table to choose a Pod. kube-proxy manages the iptables rules on the node.

This Ingress definition routes traffic for demo.example.com to a Service namedfrontend on port 80, and demo-backend.example.com to a Service namedusers on port 8080.

apiVersion: networking.k8s.io/v1
kind: Ingress
metadata:
  name: demo
spec:
  rules:
  - host: demo.example.com
    http:
      paths:
      - backend:
          service:
            name: frontend
            port:
              number: 80
  - host: demo-backend.example.com
    http:
      paths:
      - backend:
          service:
            name: users
            port:
              number: 8080

Visit GKE Ingress for Application Load Balancerfor more information.

Technical details

When you create an Ingress object, the GKE Ingress controller configures an Application Load Balancer according to the rules in the Ingress manifest and the associated Service manifests. The client sends a request to the Application Load Balancer. The load balancer is an actual proxy; it chooses a node and forwards the request to that node's NodeIP:NodePortcombination. The node uses its iptables NAT table to choose a Pod.kube-proxy manages the iptables rules on the node.

Networking security

To enhance the security of your cluster, you can limit connectivity among nodes, between Pods, and to load balancers.

Limit connectivity among nodes

Creating ingress or egress firewall rules targeting nodes in your cluster may have adverse effects. For example, applying egress deny rules to nodes in your cluster could break features such as NodePort and kubectl exec.

Limit connectivity to Pods and Services

By default, all Pods running within the same cluster can communicate freely. However, you can limit connectivity within a cluster in different ways, depending on your needs.

Limit access among Pods

You can limit access among Pods using anetwork policy. Network policy definitions allow you to restrict the ingressand egressof Pods based on an arbitrary combination of labels, IP ranges, and port numbers.

By default, there is no network policy, so all traffic among Pods in the cluster is allowed. As soon as you create the first network policy in a namespace, all other traffic is denied.

After creating a network policy, you must explicitly enable it for the cluster. For more information, seeConfiguring network policies for applications.

Limit access to an external load balancer

If your Service uses an external load balancer, traffic from any external IP address can access your Service by default. You can restrict which IP address ranges can access endpoints within your cluster, by configuring theloadBalancerSourceRanges option when configuring the Service. You can specify multiple ranges, and you can update the configuration of a running Service at any time. The kube-proxy instance running on each node configures that node'siptables rules to deny all traffic that does not match the specifiedloadBalancerSourceRanges. Additionally, when you create a LoadBalancer Service, GKE creates a corresponding VPC firewall rule to enforce these restrictions at the network level.

Limit access to an Application Load Balancer

If your service uses an Application Load Balancer, you can use a Google Cloud Armor security policy to limit which external IP addresses can access your Service and which responses to return when access is denied because of the security policy. You can configureCloud Logging to log information about these interactions.

If a Google Cloud Armor security policy is not fine-grained enough, you can enable the Identity-Aware Proxy on your endpoints to implement user-based authentication and authorization for your application. Visit thedetailed tutorial for configuring IAPfor more information.

Known issues

This section covers the known issues.

Container-enabled node unable to connect to 172.17/16 range

A node VM with containerd enabled cannot connect to a host that has an IP within 172.17/16. For more information, seeConflict with 172.17/16 IP address range.

Remaining resources from deleted GKE clusters with Private Service Connect

If you created and deleted GKE clusters with Private Service Connect before May 7, 2024, and you deleted the project containing the cluster before the cluster itself, you might have leaked associated Private Service Connect resources. These resources remain hidden and don't allow you to delete the associated subnets. If you encounter this issue, contact Google Cloud Support.

What's next