This is the multi-page printable view of this section. Click here to print.

Return to the regular view of this page.

Services, Load Balancing, and Networking

Concepts and resources behind networking in Kubernetes.

The Kubernetes network model

Every Pod in a cluster gets its own unique cluster-wide IP address. This means you do not need to explicitly create links between Pods and you almost never need to deal with mapping container ports to host ports.
This creates a clean, backwards-compatible model where Pods can be treated much like VMs or physical hosts from the perspectives of port allocation, naming, service discovery, load balancing, application configuration, and migration.

Kubernetes imposes the following fundamental requirements on any networking implementation (barring any intentional network segmentation policies):

  • pods can communicate with all other pods on any other node without NAT
  • agents on a node (e.g. system daemons, kubelet) can communicate with all pods on that node

Note: For those platforms that support Pods running in the host network (e.g. Linux), when pods are attached to the host network of a node they can still communicate with all pods on all nodes without NAT.

This model is not only less complex overall, but it is principally compatible with the desire for Kubernetes to enable low-friction porting of apps from VMs to containers. If your job previously ran in a VM, your VM had an IP and could talk to other VMs in your project. This is the same basic model.

Kubernetes IP addresses exist at the Pod scope - containers within a Pod share their network namespaces - including their IP address and MAC address. This means that containers within a Pod can all reach each other's ports on localhost. This also means that containers within a Pod must coordinate port usage, but this is no different from processes in a VM. This is called the "IP-per-pod" model.

How this is implemented is a detail of the particular container runtime in use.

It is possible to request ports on the Node itself which forward to your Pod (called host ports), but this is a very niche operation. How that forwarding is implemented is also a detail of the container runtime. The Pod itself is blind to the existence or non-existence of host ports.

Kubernetes networking addresses four concerns:

The Connecting Applications with Services tutorial lets you learn about Services and Kubernetes networking with a hands-on example.

Cluster Networking explains how to set up networking for your cluster, and also provides an overview of the technologies involved.

1 - Service

Expose an application running in your cluster behind a single outward-facing endpoint, even when the workload is split across multiple backends.
An abstract way to expose an application running on a set of Pods as a network service.

With Kubernetes you don't need to modify your application to use an unfamiliar service discovery mechanism. Kubernetes gives Pods their own IP addresses and a single DNS name for a set of Pods, and can load-balance across them.

Motivation

Kubernetes Pods are created and destroyed to match the desired state of your cluster. Pods are nonpermanent resources. If you use a Deployment to run your app, it can create and destroy Pods dynamically.

Each Pod gets its own IP address, however in a Deployment, the set of Pods running in one moment in time could be different from the set of Pods running that application a moment later.

This leads to a problem: if some set of Pods (call them "backends") provides functionality to other Pods (call them "frontends") inside your cluster, how do the frontends find out and keep track of which IP address to connect to, so that the frontend can use the backend part of the workload?

Enter Services.

Service resources

In Kubernetes, a Service is an abstraction which defines a logical set of Pods and a policy by which to access them (sometimes this pattern is called a micro-service). The set of Pods targeted by a Service is usually determined by a selector. To learn about other ways to define Service endpoints, see Services without selectors.

For example, consider a stateless image-processing backend which is running with 3 replicas. Those replicas are fungible—frontends do not care which backend they use. While the actual Pods that compose the backend set may change, the frontend clients should not need to be aware of that, nor should they need to keep track of the set of backends themselves.

The Service abstraction enables this decoupling.

Cloud-native service discovery

If you're able to use Kubernetes APIs for service discovery in your application, you can query the API server for matching EndpointSlices. Kubernetes updates the EndpointSlices for a Service whenever the set of Pods in a Service changes.

For non-native applications, Kubernetes offers ways to place a network port or load balancer in between your application and the backend Pods.

Defining a Service

A Service in Kubernetes is a REST object, similar to a Pod. Like all of the REST objects, you can POST a Service definition to the API server to create a new instance. The name of a Service object must be a valid RFC 1035 label name.

For example, suppose you have a set of Pods where each listens on TCP port 9376 and contains a label app.kubernetes.io/name=MyApp:

apiVersion: v1
kind: Service
metadata:
  name: my-service
spec:
  selector:
    app.kubernetes.io/name: MyApp
  ports:
    - protocol: TCP
      port: 80
      targetPort: 9376

This specification creates a new Service object named "my-service", which targets TCP port 9376 on any Pod with the app.kubernetes.io/name=MyApp label.

Kubernetes assigns this Service an IP address (sometimes called the "cluster IP"), which is used by the Service proxies (see Virtual IPs and service proxies below).

The controller for the Service selector continuously scans for Pods that match its selector, and then POSTs any updates to an Endpoint object also named "my-service".

Port definitions in Pods have names, and you can reference these names in the targetPort attribute of a Service. For example, we can bind the targetPort of the Service to the Pod port in the following way:

apiVersion: v1
kind: Pod
metadata:
  name: nginx
  labels:
    app.kubernetes.io/name: proxy
spec:
  containers:
  - name: nginx
    image: nginx:stable
    ports:
      - containerPort: 80
        name: http-web-svc

---
apiVersion: v1
kind: Service
metadata:
  name: nginx-service
spec:
  selector:
    app.kubernetes.io/name: proxy
  ports:
  - name: name-of-service-port
    protocol: TCP
    port: 80
    targetPort: http-web-svc

This works even if there is a mixture of Pods in the Service using a single configured name, with the same network protocol available via different port numbers. This offers a lot of flexibility for deploying and evolving your Services. For example, you can change the port numbers that Pods expose in the next version of your backend software, without breaking clients.

The default protocol for Services is TCP; you can also use any other supported protocol.

As many Services need to expose more than one port, Kubernetes supports multiple port definitions on a Service object. Each port definition can have the same protocol, or a different one.

Services without selectors

Services most commonly abstract access to Kubernetes Pods thanks to the selector, but when used with a corresponding set of EndpointSlices objects and without a selector, the Service can abstract other kinds of backends, including ones that run outside the cluster.

For example:

  • You want to have an external database cluster in production, but in your test environment you use your own databases.
  • You want to point your Service to a Service in a different Namespace or on another cluster.
  • You are migrating a workload to Kubernetes. While evaluating the approach, you run only a portion of your backends in Kubernetes.

In any of these scenarios you can define a Service without a Pod selector. For example:

apiVersion: v1
kind: Service
metadata:
  name: my-service
spec:
  ports:
    - protocol: TCP
      port: 80
      targetPort: 9376

Because this Service has no selector, the corresponding EndpointSlice (and legacy Endpoints) objects are not created automatically. You can manually map the Service to the network address and port where it's running, by adding an EndpointSlice object manually. For example:

apiVersion: discovery.k8s.io/v1
kind: EndpointSlice
metadata:
  name: my-service-1 # by convention, use the name of the Service
                     # as a prefix for the name of the EndpointSlice
  labels:
    # You should set the "kubernetes.io/service-name" label.
    # Set its value to match the name of the Service
    kubernetes.io/service-name: my-service
addressType: IPv4
ports:
  - name: '' # empty because port 9376 is not assigned as a well-known
             # port (by IANA)
    appProtocol: http
    protocol: TCP
    port: 9376
endpoints:
  - addresses:
      - "10.4.5.6" # the IP addresses in this list can appear in any order
      - "10.1.2.3"

Custom EndpointSlices

When you create an EndpointSlice object for a Service, you can use any name for the EndpointSlice. Each EndpointSlice in a namespace must have a unique name. You link an EndpointSlice to a Service by setting the kubernetes.io/service-name label on that EndpointSlice.

For an EndpointSlice that you create yourself, or in your own code, you should also pick a value to use for the endpointslice.kubernetes.io/managed-by label. If you create your own controller code to manage EndpointSlices, consider using a value similar to "my-domain.example/name-of-controller". If you are using a third party tool, use the name of the tool in all-lowercase and change spaces and other punctuation to dashes (-). If people are directly using a tool such as kubectl to manage EndpointSlices, use a name that describes this manual management, such as "staff" or "cluster-admins". You should avoid using the reserved value "controller", which identifies EndpointSlices managed by Kubernetes' own control plane.

Accessing a Service without a selector

Accessing a Service without a selector works the same as if it had a selector. In the example for a Service without a selector, traffic is routed to one of the two endpoints defined in the EndpointSlice manifest: a TCP connection to 10.1.2.3 or 10.4.5.6, on port 9376.

An ExternalName Service is a special case of Service that does not have selectors and uses DNS names instead. For more information, see the ExternalName section later in this document.

EndpointSlices

FEATURE STATE: Kubernetes v1.21 [stable]

EndpointSlices are objects that represent a subset (a slice) of the backing network endpoints for a Service.

Your Kubernetes cluster tracks how many endpoints each EndpointSlice represents. If there are so many endpoints for a Service that a threshold is reached, then Kubernetes adds another empty EndpointSlice and stores new endpoint information there. By default, Kubernetes makes a new EndpointSlice once the existing EndpointSlices all contain at least 100 endpoints. Kubernetes does not make the new EndpointSlice until an extra endpoint needs to be added.

See EndpointSlices for more information about this API.

Endpoints

In the Kubernetes API, an Endpoints (the resource kind is plural) defines a list of network endpoints, typically referenced by a Service to define which Pods the traffic can be sent to.

The EndpointSlice API is the recommended replacement for Endpoints.

Over-capacity endpoints

Kubernetes limits the number of endpoints that can fit in a single Endpoints object. When there are over 1000 backing endpoints for a Service, Kubernetes truncates the data in the Endpoints object. Because a Service can be linked with more than one EndpointSlice, the 1000 backing endpoint limit only affects the legacy Endpoints API.

In that case, Kubernetes selects at most 1000 possible backend endpoints to store into the Endpoints object, and sets an annotation on the Endpoints: endpoints.kubernetes.io/over-capacity: truncated. The control plane also removes that annotation if the number of backend Pods drops below 1000.

Traffic is still sent to backends, but any load balancing mechanism that relies on the legacy Endpoints API only sends traffic to at most 1000 of the available backing endpoints.

The same API limit means that you cannot manually update an Endpoints to have more than 1000 endpoints.

Application protocol

FEATURE STATE: Kubernetes v1.20 [stable]

The appProtocol field provides a way to specify an application protocol for each Service port. The value of this field is mirrored by the corresponding Endpoints and EndpointSlice objects.

This field follows standard Kubernetes label syntax. Values should either be IANA standard service names or domain prefixed names such as mycompany.com/my-custom-protocol.

Virtual IPs and service proxies

Every node in a Kubernetes cluster runs a kube-proxy. kube-proxy is responsible for implementing a form of virtual IP for Services of type other than ExternalName.

Why not use round-robin DNS?

A question that pops up every now and then is why Kubernetes relies on proxying to forward inbound traffic to backends. What about other approaches? For example, would it be possible to configure DNS records that have multiple A values (or AAAA for IPv6), and rely on round-robin name resolution?

There are a few reasons for using proxying for Services:

  • There is a long history of DNS implementations not respecting record TTLs, and caching the results of name lookups after they should have expired.
  • Some apps do DNS lookups only once and cache the results indefinitely.
  • Even if apps and libraries did proper re-resolution, the low or zero TTLs on the DNS records could impose a high load on DNS that then becomes difficult to manage.

Later in this page you can read about how various kube-proxy implementations work. Overall, you should note that, when running kube-proxy, kernel level rules may be modified (for example, iptables rules might get created), which won't get cleaned up, in some cases until you reboot. Thus, running kube-proxy is something that should only be done by an administrator which understands the consequences of having a low level, privileged network proxying service on a computer. Although the kube-proxy executable supports a cleanup function, this function is not an official feature and thus is only available to use as-is.

Configuration

Note that the kube-proxy starts up in different modes, which are determined by its configuration.

  • The kube-proxy's configuration is done via a ConfigMap, and the ConfigMap for kube-proxy effectively deprecates the behavior for almost all of the flags for the kube-proxy.
  • The ConfigMap for the kube-proxy does not support live reloading of configuration.
  • The ConfigMap parameters for the kube-proxy cannot all be validated and verified on startup. For example, if your operating system doesn't allow you to run iptables commands, the standard kernel kube-proxy implementation will not work. Likewise, if you have an operating system which doesn't support netsh, it will not run in Windows userspace mode.

User space proxy mode

In this (legacy) mode, kube-proxy watches the Kubernetes control plane for the addition and removal of Service and Endpoint objects. For each Service it opens a port (randomly chosen) on the local node. Any connections to this "proxy port" are proxied to one of the Service's backend Pods (as reported via Endpoints). kube-proxy takes the SessionAffinity setting of the Service into account when deciding which backend Pod to use.

Lastly, the user-space proxy installs iptables rules which capture traffic to the Service's clusterIP (which is virtual) and port. The rules redirect that traffic to the proxy port which proxies the backend Pod.

By default, kube-proxy in userspace mode chooses a backend via a round-robin algorithm.

Services overview diagram for userspace proxy

iptables proxy mode

In this mode, kube-proxy watches the Kubernetes control plane for the addition and removal of Service and Endpoint objects. For each Service, it installs iptables rules, which capture traffic to the Service's clusterIP and port, and redirect that traffic to one of the Service's backend sets. For each Endpoint object, it installs iptables rules which select a backend Pod.

By default, kube-proxy in iptables mode chooses a backend at random.

Using iptables to handle traffic has a lower system overhead, because traffic is handled by Linux netfilter without the need to switch between userspace and the kernel space. This approach is also likely to be more reliable.

If kube-proxy is running in iptables mode and the first Pod that's selected does not respond, the connection fails. This is different from userspace mode: in that scenario, kube-proxy would detect that the connection to the first Pod had failed and would automatically retry with a different backend Pod.

You can use Pod readiness probes to verify that backend Pods are working OK, so that kube-proxy in iptables mode only sees backends that test out as healthy. Doing this means you avoid having traffic sent via kube-proxy to a Pod that's known to have failed.

Services overview diagram for iptables proxy

IPVS proxy mode

FEATURE STATE: Kubernetes v1.11 [stable]

In ipvs mode, kube-proxy watches Kubernetes Services and Endpoints, calls netlink interface to create IPVS rules accordingly and synchronizes IPVS rules with Kubernetes Services and Endpoints periodically. This control loop ensures that IPVS status matches the desired state. When accessing a Service, IPVS directs traffic to one of the backend Pods.

The IPVS proxy mode is based on netfilter hook function that is similar to iptables mode, but uses a hash table as the underlying data structure and works in the kernel space. That means kube-proxy in IPVS mode redirects traffic with lower latency than kube-proxy in iptables mode, with much better performance when synchronizing proxy rules. Compared to the other proxy modes, IPVS mode also supports a higher throughput of network traffic.

IPVS provides more options for balancing traffic to backend Pods; these are:

  • rr: round-robin
  • lc: least connection (smallest number of open connections)
  • dh: destination hashing
  • sh: source hashing
  • sed: shortest expected delay
  • nq: never queue

Services overview diagram for IPVS proxy

In these proxy models, the traffic bound for the Service's IP:Port is proxied to an appropriate backend without the clients knowing anything about Kubernetes or Services or Pods.

If you want to make sure that connections from a particular client are passed to the same Pod each time, you can select the session affinity based on the client's IP addresses by setting service.spec.sessionAffinity to "ClientIP" (the default is "None"). You can also set the maximum session sticky time by setting service.spec.sessionAffinityConfig.clientIP.timeoutSeconds appropriately. (the default value is 10800, which works out to be 3 hours).

Multi-Port Services

For some Services, you need to expose more than one port. Kubernetes lets you configure multiple port definitions on a Service object. When using multiple ports for a Service, you must give all of your ports names so that these are unambiguous. For example:

apiVersion: v1
kind: Service
metadata:
  name: my-service
spec:
  selector:
    app.kubernetes.io/name: MyApp
  ports:
    - name: http
      protocol: TCP
      port: 80
      targetPort: 9376
    - name: https
      protocol: TCP
      port: 443
      targetPort: 9377

Choosing your own IP address

You can specify your own cluster IP address as part of a Service creation request. To do this, set the .spec.clusterIP field. For example, if you already have an existing DNS entry that you wish to reuse, or legacy systems that are configured for a specific IP address and difficult to re-configure.

The IP address that you choose must be a valid IPv4 or IPv6 address from within the service-cluster-ip-range CIDR range that is configured for the API server. If you try to create a Service with an invalid clusterIP address value, the API server will return a 422 HTTP status code to indicate that there's a problem.

Traffic policies

External traffic policy

You can set the spec.externalTrafficPolicy field to control how traffic from external sources is routed. Valid values are Cluster and Local. Set the field to Cluster to route external traffic to all ready endpoints and Local to only route to ready node-local endpoints. If the traffic policy is Local and there are no node-local endpoints, the kube-proxy does not forward any traffic for the relevant Service.

Internal traffic policy

FEATURE STATE: Kubernetes v1.22 [beta]

You can set the spec.internalTrafficPolicy field to control how traffic from internal sources is routed. Valid values are Cluster and Local. Set the field to Cluster to route internal traffic to all ready endpoints and Local to only route to ready node-local endpoints. If the traffic policy is Local and there are no node-local endpoints, traffic is dropped by kube-proxy.

Discovering services

Kubernetes supports 2 primary modes of finding a Service - environment variables and DNS.

Environment variables

When a Pod is run on a Node, the kubelet adds a set of environment variables for each active Service. It adds {SVCNAME}_SERVICE_HOST and {SVCNAME}_SERVICE_PORT variables, where the Service name is upper-cased and dashes are converted to underscores. It also supports variables (see makeLinkVariables) that are compatible with Docker Engine's "legacy container links" feature.

For example, the Service redis-primary which exposes TCP port 6379 and has been allocated cluster IP address 10.0.0.11, produces the following environment variables:

REDIS_PRIMARY_SERVICE_HOST=10.0.0.11
REDIS_PRIMARY_SERVICE_PORT=6379
REDIS_PRIMARY_PORT=tcp://10.0.0.11:6379
REDIS_PRIMARY_PORT_6379_TCP=tcp://10.0.0.11:6379
REDIS_PRIMARY_PORT_6379_TCP_PROTO=tcp
REDIS_PRIMARY_PORT_6379_TCP_PORT=6379
REDIS_PRIMARY_PORT_6379_TCP_ADDR=10.0.0.11

DNS

You can (and almost always should) set up a DNS service for your Kubernetes cluster using an add-on.

A cluster-aware DNS server, such as CoreDNS, watches the Kubernetes API for new Services and creates a set of DNS records for each one. If DNS has been enabled throughout your cluster then all Pods should automatically be able to resolve Services by their DNS name.

For example, if you have a Service called my-service in a Kubernetes namespace my-ns, the control plane and the DNS Service acting together create a DNS record for my-service.my-ns. Pods in the my-ns namespace should be able to find the service by doing a name lookup for my-service (my-service.my-ns would also work).

Pods in other namespaces must qualify the name as my-service.my-ns. These names will resolve to the cluster IP assigned for the Service.

Kubernetes also supports DNS SRV (Service) records for named ports. If the my-service.my-ns Service has a port named http with the protocol set to TCP, you can do a DNS SRV query for _http._tcp.my-service.my-ns to discover the port number for http, as well as the IP address.

The Kubernetes DNS server is the only way to access ExternalName Services. You can find more information about ExternalName resolution in DNS Pods and Services.

Headless Services

Sometimes you don't need load-balancing and a single Service IP. In this case, you can create what are termed "headless" Services, by explicitly specifying "None" for the cluster IP (.spec.clusterIP).

You can use a headless Service to interface with other service discovery mechanisms, without being tied to Kubernetes' implementation.

For headless Services, a cluster IP is not allocated, kube-proxy does not handle these Services, and there is no load balancing or proxying done by the platform for them. How DNS is automatically configured depends on whether the Service has selectors defined:

With selectors

For headless Services that define selectors, the Kubernetes control plane creates EndpointSlice objects in the Kubernetes API, and modifies the DNS configuration to return A or AAAA records (IPv4 or IPv6 addresses) that point directly to the Pods backing the Service.

Without selectors

For headless Services that do not define selectors, the control plane does not create EndpointSlice objects. However, the DNS system looks for and configures either:

  • DNS CNAME records for type: ExternalName Services.
  • DNS A / AAAA records for all IP addresses of the Service's ready endpoints, for all Service types other than ExternalName.
    • For IPv4 endpoints, the DNS system creates A records.
    • For IPv6 endpoints, the DNS system creates AAAA records.

Publishing Services (ServiceTypes)

For some parts of your application (for example, frontends) you may want to expose a Service onto an external IP address, that's outside of your cluster.

Kubernetes ServiceTypes allow you to specify what kind of Service you want.

Type values and their behaviors are:

  • ClusterIP: Exposes the Service on a cluster-internal IP. Choosing this value makes the Service only reachable from within the cluster. This is the default that is used if you don't explicitly specify a type for a Service.
  • NodePort: Exposes the Service on each Node's IP at a static port (the NodePort). To make the node port available, Kubernetes sets up a cluster IP address, the same as if you had requested a Service of type: ClusterIP.
  • LoadBalancer: Exposes the Service externally using a cloud provider's load balancer.
  • ExternalName: Maps the Service to the contents of the externalName field (e.g. foo.bar.example.com), by returning a CNAME record with its value. No proxying of any kind is set up.

You can also use Ingress to expose your Service. Ingress is not a Service type, but it acts as the entry point for your cluster. It lets you consolidate your routing rules into a single resource as it can expose multiple services under the same IP address.

Type NodePort

If you set the type field to NodePort, the Kubernetes control plane allocates a port from a range specified by --service-node-port-range flag (default: 30000-32767). Each node proxies that port (the same port number on every Node) into your Service. Your Service reports the allocated port in its .spec.ports[*].nodePort field.

Using a NodePort gives you the freedom to set up your own load balancing solution, to configure environments that are not fully supported by Kubernetes, or even to expose one or more nodes' IP addresses directly.

For a node port Service, Kubernetes additionally allocates a port (TCP, UDP or SCTP to match the protocol of the Service). Every node in the cluster configures itself to listen on that assigned port and to forward traffic to one of the ready endpoints associated with that Service. You'll be able to contact the type: NodePort Service, from outside the cluster, by connecting to any node using the appropriate protocol (for example: TCP), and the appropriate port (as assigned to that Service).

Choosing your own port

If you want a specific port number, you can specify a value in the nodePort field. The control plane will either allocate you that port or report that the API transaction failed. This means that you need to take care of possible port collisions yourself. You also have to use a valid port number, one that's inside the range configured for NodePort use.

Here is an example manifest for a Service of type: NodePort that specifies a NodePort value (30007, in this example).

apiVersion: v1
kind: Service
metadata:
  name: my-service
spec:
  type: NodePort
  selector:
    app.kubernetes.io/name: MyApp
  ports:
      # By default and for convenience, the `targetPort` is set to the same value as the `port` field.
    - port: 80
      targetPort: 80
      # Optional field
      # By default and for convenience, the Kubernetes control plane will allocate a port from a range (default: 30000-32767)
      nodePort: 30007

Custom IP address configuration for type: NodePort Services

You can set up nodes in your cluster to use a particular IP address for serving node port services. You might want to do this if each node is connected to multiple networks (for example: one network for application traffic, and another network for traffic between nodes and the control plane).

If you want to specify particular IP address(es) to proxy the port, you can set the --nodeport-addresses flag for kube-proxy or the equivalent nodePortAddresses field of the kube-proxy configuration file to particular IP block(s).

This flag takes a comma-delimited list of IP blocks (e.g. 10.0.0.0/8, 192.0.2.0/25) to specify IP address ranges that kube-proxy should consider as local to this node.

For example, if you start kube-proxy with the --nodeport-addresses=127.0.0.0/8 flag, kube-proxy only selects the loopback interface for NodePort Services. The default for --nodeport-addresses is an empty list. This means that kube-proxy should consider all available network interfaces for NodePort. (That's also compatible with earlier Kubernetes releases.)

Type LoadBalancer

On cloud providers which support external load balancers, setting the type field to LoadBalancer provisions a load balancer for your Service. The actual creation of the load balancer happens asynchronously, and information about the provisioned balancer is published in the Service's .status.loadBalancer field. For example:

apiVersion: v1
kind: Service
metadata:
  name: my-service
spec:
  selector:
    app.kubernetes.io/name: MyApp
  ports:
    - protocol: TCP
      port: 80
      targetPort: 9376
  clusterIP: 10.0.171.239
  type: LoadBalancer
status:
  loadBalancer:
    ingress:
    - ip: 192.0.2.127

Traffic from the external load balancer is directed at the backend Pods. The cloud provider decides how it is load balanced.

Some cloud providers allow you to specify the loadBalancerIP. In those cases, the load-balancer is created with the user-specified loadBalancerIP. If the loadBalancerIP field is not specified, the loadBalancer is set up with an ephemeral IP address. If you specify a loadBalancerIP but your cloud provider does not support the feature, the loadbalancerIP field that you set is ignored.

To implement a Service of type: LoadBalancer, Kubernetes typically starts off by making the changes that are equivalent to you requesting a Service of type: NodePort. The cloud-controller-manager component then configures the external load balancer to forward traffic to that assigned node port.

As an alpha feature, you can configure a load balanced Service to omit assigning a node port, provided that the cloud provider implementation supports this.

Load balancers with mixed protocol types

FEATURE STATE: Kubernetes v1.24 [beta]

By default, for LoadBalancer type of Services, when there is more than one port defined, all ports must have the same protocol, and the protocol must be one which is supported by the cloud provider.

The feature gate MixedProtocolLBService (enabled by default for the kube-apiserver as of v1.24) allows the use of different protocols for LoadBalancer type of Services, when there is more than one port defined.

Disabling load balancer NodePort allocation

FEATURE STATE: Kubernetes v1.24 [stable]

You can optionally disable node port allocation for a Service of type=LoadBalancer, by setting the field spec.allocateLoadBalancerNodePorts to false. This should only be used for load balancer implementations that route traffic directly to pods as opposed to using node ports. By default, spec.allocateLoadBalancerNodePorts is true and type LoadBalancer Services will continue to allocate node ports. If spec.allocateLoadBalancerNodePorts is set to false on an existing Service with allocated node ports, those node ports will not be de-allocated automatically. You must explicitly remove the nodePorts entry in every Service port to de-allocate those node ports.

Specifying class of load balancer implementation

FEATURE STATE: Kubernetes v1.24 [stable]

spec.loadBalancerClass enables you to use a load balancer implementation other than the cloud provider default. By default, spec.loadBalancerClass is nil and a LoadBalancer type of Service uses the cloud provider's default load balancer implementation if the cluster is configured with a cloud provider using the --cloud-provider component flag. If spec.loadBalancerClass is specified, it is assumed that a load balancer implementation that matches the specified class is watching for Services. Any default load balancer implementation (for example, the one provided by the cloud provider) will ignore Services that have this field set. spec.loadBalancerClass can be set on a Service of type LoadBalancer only. Once set, it cannot be changed. The value of spec.loadBalancerClass must be a label-style identifier, with an optional prefix such as "internal-vip" or "example.com/internal-vip". Unprefixed names are reserved for end-users.

Internal load balancer

In a mixed environment it is sometimes necessary to route traffic from Services inside the same (virtual) network address block.

In a split-horizon DNS environment you would need two Services to be able to route both external and internal traffic to your endpoints.

To set an internal load balancer, add one of the following annotations to your Service depending on the cloud Service provider you're using.

Select one of the tabs.

[...]
metadata:
    name: my-service
    annotations:
        cloud.google.com/load-balancer-type: "Internal"
[...]

[...]
metadata:
    name: my-service
    annotations:
        service.beta.kubernetes.io/aws-load-balancer-internal: "true"
[...]

[...]
metadata:
    name: my-service
    annotations:
        service.beta.kubernetes.io/azure-load-balancer-internal: "true"
[...]

[...]
metadata:
    name: my-service
    annotations:
        service.kubernetes.io/ibm-load-balancer-cloud-provider-ip-type: "private"
[...]

[...]
metadata:
    name: my-service
    annotations:
        service.beta.kubernetes.io/openstack-internal-load-balancer: "true"
[...]

[...]
metadata:
    name: my-service
    annotations:
        service.beta.kubernetes.io/cce-load-balancer-internal-vpc: "true"
[...]

[...]
metadata:
  annotations:
    service.kubernetes.io/qcloud-loadbalancer-internal-subnetid: subnet-xxxxx
[...]

[...]
metadata:
  annotations:
    service.beta.kubernetes.io/alibaba-cloud-loadbalancer-address-type: "intranet"
[...]

[...]
metadata:
    name: my-service
    annotations:
        service.beta.kubernetes.io/oci-load-balancer-internal: true
[...]

TLS support on AWS

For partial TLS / SSL support on clusters running on AWS, you can add three annotations to a LoadBalancer service:

metadata:
  name: my-service
  annotations:
    service.beta.kubernetes.io/aws-load-balancer-ssl-cert: arn:aws:acm:us-east-1:123456789012:certificate/12345678-1234-1234-1234-123456789012

The first specifies the ARN of the certificate to use. It can be either a certificate from a third party issuer that was uploaded to IAM or one created within AWS Certificate Manager.

metadata:
  name: my-service
  annotations:
    service.beta.kubernetes.io/aws-load-balancer-backend-protocol: (https|http|ssl|tcp)

The second annotation specifies which protocol a Pod speaks. For HTTPS and SSL, the ELB expects the Pod to authenticate itself over the encrypted connection, using a certificate.

HTTP and HTTPS selects layer 7 proxying: the ELB terminates the connection with the user, parses headers, and injects the X-Forwarded-For header with the user's IP address (Pods only see the IP address of the ELB at the other end of its connection) when forwarding requests.

TCP and SSL selects layer 4 proxying: the ELB forwards traffic without modifying the headers.

In a mixed-use environment where some ports are secured and others are left unencrypted, you can use the following annotations:

    metadata:
      name: my-service
      annotations:
        service.beta.kubernetes.io/aws-load-balancer-backend-protocol: http
        service.beta.kubernetes.io/aws-load-balancer-ssl-ports: "443,8443"

In the above example, if the Service contained three ports, 80, 443, and 8443, then 443 and 8443 would use the SSL certificate, but 80 would be proxied HTTP.

From Kubernetes v1.9 onwards you can use predefined AWS SSL policies with HTTPS or SSL listeners for your Services. To see which policies are available for use, you can use the aws command line tool:

aws elb describe-load-balancer-policies --query 'PolicyDescriptions[].PolicyName'

You can then specify any one of those policies using the "service.beta.kubernetes.io/aws-load-balancer-ssl-negotiation-policy" annotation; for example:

    metadata:
      name: my-service
      annotations:
        service.beta.kubernetes.io/aws-load-balancer-ssl-negotiation-policy: "ELBSecurityPolicy-TLS-1-2-2017-01"

PROXY protocol support on AWS

To enable PROXY protocol support for clusters running on AWS, you can use the following service annotation:

    metadata:
      name: my-service
      annotations:
        service.beta.kubernetes.io/aws-load-balancer-proxy-protocol: "*"

Since version 1.3.0, the use of this annotation applies to all ports proxied by the ELB and cannot be configured otherwise.

ELB Access Logs on AWS

There are several annotations to manage access logs for ELB Services on AWS.

The annotation service.beta.kubernetes.io/aws-load-balancer-access-log-enabled controls whether access logs are enabled.

The annotation service.beta.kubernetes.io/aws-load-balancer-access-log-emit-interval controls the interval in minutes for publishing the access logs. You can specify an interval of either 5 or 60 minutes.

The annotation service.beta.kubernetes.io/aws-load-balancer-access-log-s3-bucket-name controls the name of the Amazon S3 bucket where load balancer access logs are stored.

The annotation service.beta.kubernetes.io/aws-load-balancer-access-log-s3-bucket-prefix specifies the logical hierarchy you created for your Amazon S3 bucket.

    metadata:
      name: my-service
      annotations:
        # Specifies whether access logs are enabled for the load balancer
        service.beta.kubernetes.io/aws-load-balancer-access-log-enabled: "true"

        # The interval for publishing the access logs. You can specify an interval of either 5 or 60 (minutes).
        service.beta.kubernetes.io/aws-load-balancer-access-log-emit-interval: "60"

        # The name of the Amazon S3 bucket where the access logs are stored
        service.beta.kubernetes.io/aws-load-balancer-access-log-s3-bucket-name: "my-bucket"

        # The logical hierarchy you created for your Amazon S3 bucket, for example `my-bucket-prefix/prod`
        service.beta.kubernetes.io/aws-load-balancer-access-log-s3-bucket-prefix: "my-bucket-prefix/prod"

Connection Draining on AWS

Connection draining for Classic ELBs can be managed with the annotation service.beta.kubernetes.io/aws-load-balancer-connection-draining-enabled set to the value of "true". The annotation service.beta.kubernetes.io/aws-load-balancer-connection-draining-timeout can also be used to set maximum time, in seconds, to keep the existing connections open before deregistering the instances.

    metadata:
      name: my-service
      annotations:
        service.beta.kubernetes.io/aws-load-balancer-connection-draining-enabled: "true"
        service.beta.kubernetes.io/aws-load-balancer-connection-draining-timeout: "60"

Other ELB annotations

There are other annotations to manage Classic Elastic Load Balancers that are described below.

    metadata:
      name: my-service
      annotations:
        # The time, in seconds, that the connection is allowed to be idle (no data has been sent
        # over the connection) before it is closed by the load balancer
        service.beta.kubernetes.io/aws-load-balancer-connection-idle-timeout: "60"

        # Specifies whether cross-zone load balancing is enabled for the load balancer
        service.beta.kubernetes.io/aws-load-balancer-cross-zone-load-balancing-enabled: "true"

        # A comma-separated list of key-value pairs which will be recorded as
        # additional tags in the ELB.
        service.beta.kubernetes.io/aws-load-balancer-additional-resource-tags: "environment=prod,owner=devops"

        # The number of successive successful health checks required for a backend to
        # be considered healthy for traffic. Defaults to 2, must be between 2 and 10
        service.beta.kubernetes.io/aws-load-balancer-healthcheck-healthy-threshold: ""

        # The number of unsuccessful health checks required for a backend to be
        # considered unhealthy for traffic. Defaults to 6, must be between 2 and 10
        service.beta.kubernetes.io/aws-load-balancer-healthcheck-unhealthy-threshold: "3"

        # The approximate interval, in seconds, between health checks of an
        # individual instance. Defaults to 10, must be between 5 and 300
        service.beta.kubernetes.io/aws-load-balancer-healthcheck-interval: "20"

        # The amount of time, in seconds, during which no response means a failed
        # health check. This value must be less than the service.beta.kubernetes.io/aws-load-balancer-healthcheck-interval
        # value. Defaults to 5, must be between 2 and 60
        service.beta.kubernetes.io/aws-load-balancer-healthcheck-timeout: "5"

        # A list of existing security groups to be configured on the ELB created. Unlike the annotation
        # service.beta.kubernetes.io/aws-load-balancer-extra-security-groups, this replaces all other
        # security groups previously assigned to the ELB and also overrides the creation
        # of a uniquely generated security group for this ELB.
        # The first security group ID on this list is used as a source to permit incoming traffic to
        # target worker nodes (service traffic and health checks).
        # If multiple ELBs are configured with the same security group ID, only a single permit line
        # will be added to the worker node security groups, that means if you delete any
        # of those ELBs it will remove the single permit line and block access for all ELBs that shared the same security group ID.
        # This can cause a cross-service outage if not used properly
        service.beta.kubernetes.io/aws-load-balancer-security-groups: "sg-53fae93f"

        # A list of additional security groups to be added to the created ELB, this leaves the uniquely
        # generated security group in place, this ensures that every ELB
        # has a unique security group ID and a matching permit line to allow traffic to the target worker nodes
        # (service traffic and health checks).
        # Security groups defined here can be shared between services.
        service.beta.kubernetes.io/aws-load-balancer-extra-security-groups: "sg-53fae93f,sg-42efd82e"

        # A comma separated list of key-value pairs which are used
        # to select the target nodes for the load balancer
        service.beta.kubernetes.io/aws-load-balancer-target-node-labels: "ingress-gw,gw-name=public-api"

Network Load Balancer support on AWS

FEATURE STATE: Kubernetes v1.15 [beta]

To use a Network Load Balancer on AWS, use the annotation service.beta.kubernetes.io/aws-load-balancer-type with the value set to nlb.

    metadata:
      name: my-service
      annotations:
        service.beta.kubernetes.io/aws-load-balancer-type: "nlb"

Unlike Classic Elastic Load Balancers, Network Load Balancers (NLBs) forward the client's IP address through to the node. If a Service's .spec.externalTrafficPolicy is set to Cluster, the client's IP address is not propagated to the end Pods.

By setting .spec.externalTrafficPolicy to Local, the client IP addresses is propagated to the end Pods, but this could result in uneven distribution of traffic. Nodes without any Pods for a particular LoadBalancer Service will fail the NLB Target Group's health check on the auto-assigned .spec.healthCheckNodePort and not receive any traffic.

In order to achieve even traffic, either use a DaemonSet or specify a pod anti-affinity to not locate on the same node.

You can also use NLB Services with the internal load balancer annotation.

In order for client traffic to reach instances behind an NLB, the Node security groups are modified with the following IP rules:

Rule Protocol Port(s) IpRange(s) IpRange Description
Health Check TCP NodePort(s) (.spec.healthCheckNodePort for .spec.externalTrafficPolicy = Local) Subnet CIDR kubernetes.io/rule/nlb/health=<loadBalancerName>
Client Traffic TCP NodePort(s) .spec.loadBalancerSourceRanges (defaults to 0.0.0.0/0) kubernetes.io/rule/nlb/client=<loadBalancerName>
MTU Discovery ICMP 3,4 .spec.loadBalancerSourceRanges (defaults to 0.0.0.0/0) kubernetes.io/rule/nlb/mtu=<loadBalancerName>

In order to limit which client IP's can access the Network Load Balancer, specify loadBalancerSourceRanges.

spec:
  loadBalancerSourceRanges:
    - "143.231.0.0/16"

Further documentation on annotations for Elastic IPs and other common use-cases may be found in the AWS Load Balancer Controller documentation.

Other CLB annotations on Tencent Kubernetes Engine (TKE)

There are other annotations for managing Cloud Load Balancers on TKE as shown below.

    metadata:
      name: my-service
      annotations:
        # Bind Loadbalancers with specified nodes
        service.kubernetes.io/qcloud-loadbalancer-backends-label: key in (value1, value2)

        # ID of an existing load balancer
        service.kubernetes.io/tke-existed-lbid:lb-6swtxxxx

        # Custom parameters for the load balancer (LB), does not support modification of LB type yet
        service.kubernetes.io/service.extensiveParameters: ""

        # Custom parameters for the LB listener
        service.kubernetes.io/service.listenerParameters: ""

        # Specifies the type of Load balancer;
        # valid values: classic (Classic Cloud Load Balancer) or application (Application Cloud Load Balancer)
        service.kubernetes.io/loadbalance-type: xxxxx

        # Specifies the public network bandwidth billing method;
        # valid values: TRAFFIC_POSTPAID_BY_HOUR(bill-by-traffic) and BANDWIDTH_POSTPAID_BY_HOUR (bill-by-bandwidth).
        service.kubernetes.io/qcloud-loadbalancer-internet-charge-type: xxxxxx

        # Specifies the bandwidth value (value range: [1,2000] Mbps).
        service.kubernetes.io/qcloud-loadbalancer-internet-max-bandwidth-out: "10"

        # When this annotation is set,the loadbalancers will only register nodes
        # with pod running on it, otherwise all nodes will be registered.
        service.kubernetes.io/local-svc-only-bind-node-with-pod: true

Type ExternalName

Services of type ExternalName map a Service to a DNS name, not to a typical selector such as my-service or cassandra. You specify these Services with the spec.externalName parameter.

This Service definition, for example, maps the my-service Service in the prod namespace to my.database.example.com:

apiVersion: v1
kind: Service
metadata:
  name: my-service
  namespace: prod
spec:
  type: ExternalName
  externalName: my.database.example.com

When looking up the host my-service.prod.svc.cluster.local, the cluster DNS Service returns a CNAME record with the value my.database.example.com. Accessing my-service works in the same way as other Services but with the crucial difference that redirection happens at the DNS level rather than via proxying or forwarding. Should you later decide to move your database into your cluster, you can start its Pods, add appropriate selectors or endpoints, and change the Service's type.

External IPs

If there are external IPs that route to one or more cluster nodes, Kubernetes Services can be exposed on those externalIPs. Traffic that ingresses into the cluster with the external IP (as destination IP), on the Service port, will be routed to one of the Service endpoints. externalIPs are not managed by Kubernetes and are the responsibility of the cluster administrator.

In the Service spec, externalIPs can be specified along with any of the ServiceTypes. In the example below, "my-service" can be accessed by clients on "80.11.12.10:80" (externalIP:port)

apiVersion: v1
kind: Service
metadata:
  name: my-service
spec:
  selector:
    app.kubernetes.io/name: MyApp
  ports:
    - name: http
      protocol: TCP
      port: 80
      targetPort: 9376
  externalIPs:
    - 80.11.12.10

Shortcomings

Using the userspace proxy for VIPs works at small to medium scale, but will not scale to very large clusters with thousands of Services. The original design proposal for portals has more details on this.

Using the userspace proxy obscures the source IP address of a packet accessing a Service. This makes some kinds of network filtering (firewalling) impossible. The iptables proxy mode does not obscure in-cluster source IPs, but it does still impact clients coming through a load balancer or node-port.

The Type field is designed as nested functionality - each level adds to the previous. This is not strictly required on all cloud providers (e.g. Google Compute Engine does not need to allocate a NodePort to make LoadBalancer work, but AWS does) but the Kubernetes API design for Service requires it anyway.

Virtual IP implementation

The previous information should be sufficient for many people who want to use Services. However, there is a lot going on behind the scenes that may be worth understanding.

Avoiding collisions

One of the primary philosophies of Kubernetes is that you should not be exposed to situations that could cause your actions to fail through no fault of your own. For the design of the Service resource, this means not making you choose your own port number if that choice might collide with someone else's choice. That is an isolation failure.

In order to allow you to choose a port number for your Services, we must ensure that no two Services can collide. Kubernetes does that by allocating each Service its own IP address from within the service-cluster-ip-range CIDR range that is configured for the API server.

To ensure each Service receives a unique IP, an internal allocator atomically updates a global allocation map in etcd prior to creating each Service. The map object must exist in the registry for Services to get IP address assignments, otherwise creations will fail with a message indicating an IP address could not be allocated.

In the control plane, a background controller is responsible for creating that map (needed to support migrating from older versions of Kubernetes that used in-memory locking). Kubernetes also uses controllers to check for invalid assignments (e.g. due to administrator intervention) and for cleaning up allocated IP addresses that are no longer used by any Services.

IP address ranges for type: ClusterIP Services

FEATURE STATE: Kubernetes v1.25 [beta]
However, there is a problem with this ClusterIP allocation strategy, because a user can also choose their own address for the service. This could result in a conflict if the internal allocator selects the same IP address for another Service.

The ServiceIPStaticSubrange feature gate is enabled by default in v1.25 and later, using an allocation strategy that divides the ClusterIP range into two bands, based on the size of the configured service-cluster-ip-range by using the following formula min(max(16, cidrSize / 16), 256), described as never less than 16 or more than 256, with a graduated step function between them. Dynamic IP allocations will be preferentially chosen from the upper band, reducing risks of conflicts with the IPs assigned from the lower band. This allows users to use the lower band of the service-cluster-ip-range for their Services with static IPs assigned with a very low risk of running into conflicts.

Service IP addresses

Unlike Pod IP addresses, which actually route to a fixed destination, Service IPs are not actually answered by a single host. Instead, kube-proxy uses iptables (packet processing logic in Linux) to define virtual IP addresses which are transparently redirected as needed. When clients connect to the VIP, their traffic is automatically transported to an appropriate endpoint. The environment variables and DNS for Services are actually populated in terms of the Service's virtual IP address (and port).

kube-proxy supports three proxy modes—userspace, iptables and IPVS—which each operate slightly differently.

Userspace

As an example, consider the image processing application described above. When the backend Service is created, the Kubernetes master assigns a virtual IP address, for example 10.0.0.1. Assuming the Service port is 1234, the Service is observed by all of the kube-proxy instances in the cluster. When a proxy sees a new Service, it opens a new random port, establishes an iptables redirect from the virtual IP address to this new port, and starts accepting connections on it.

When a client connects to the Service's virtual IP address, the iptables rule kicks in, and redirects the packets to the proxy's own port. The "Service proxy" chooses a backend, and starts proxying traffic from the client to the backend.

This means that Service owners can choose any port they want without risk of collision. Clients can connect to an IP and port, without being aware of which Pods they are actually accessing.

iptables

Again, consider the image processing application described above. When the backend Service is created, the Kubernetes control plane assigns a virtual IP address, for example 10.0.0.1. Assuming the Service port is 1234, the Service is observed by all of the kube-proxy instances in the cluster. When a proxy sees a new Service, it installs a series of iptables rules which redirect from the virtual IP address to per-Service rules. The per-Service rules link to per-Endpoint rules which redirect traffic (using destination NAT) to the backends.

When a client connects to the Service's virtual IP address the iptables rule kicks in. A backend is chosen (either based on session affinity or randomly) and packets are redirected to the backend. Unlike the userspace proxy, packets are never copied to userspace, the kube-proxy does not have to be running for the virtual IP address to work, and Nodes see traffic arriving from the unaltered client IP address.

This same basic flow executes when traffic comes in through a node-port or through a load-balancer, though in those cases the client IP does get altered.

IPVS

iptables operations slow down dramatically in large scale cluster e.g. 10,000 Services. IPVS is designed for load balancing and based on in-kernel hash tables. So you can achieve performance consistency in large number of Services from IPVS-based kube-proxy. Meanwhile, IPVS-based kube-proxy has more sophisticated load balancing algorithms (least conns, locality, weighted, persistence).

API Object

Service is a top-level resource in the Kubernetes REST API. You can find more details about the Service API object.

Supported protocols

TCP

You can use TCP for any kind of Service, and it's the default network protocol.

UDP

You can use UDP for most Services. For type=LoadBalancer Services, UDP support depends on the cloud provider offering this facility.

SCTP

FEATURE STATE: Kubernetes v1.20 [stable]

When using a network plugin that supports SCTP traffic, you can use SCTP for most Services. For type=LoadBalancer Services, SCTP support depends on the cloud provider offering this facility. (Most do not).

Warnings

Support for multihomed SCTP associations
Windows
Userspace kube-proxy

HTTP

If your cloud provider supports it, you can use a Service in LoadBalancer mode to set up external HTTP / HTTPS reverse proxying, forwarded to the Endpoints of the Service.

PROXY protocol

If your cloud provider supports it, you can use a Service in LoadBalancer mode to configure a load balancer outside of Kubernetes itself, that will forward connections prefixed with PROXY protocol.

The load balancer will send an initial series of octets describing the incoming connection, similar to this example

PROXY TCP4 192.0.2.202 10.0.42.7 12345 7\r\n

followed by the data from the client.

What's next

2 - Ingress

Make your HTTP (or HTTPS) network service available using a protocol-aware configuration mechanism, that understands web concepts like URIs, hostnames, paths, and more. The Ingress concept lets you map traffic to different backends based on rules you define via the Kubernetes API.

FEATURE STATE: Kubernetes v1.19 [stable]

An API object that manages external access to the services in a cluster, typically HTTP.

Ingress may provide load balancing, SSL termination and name-based virtual hosting.

Terminology

For clarity, this guide defines the following terms:

  • Node: A worker machine in Kubernetes, part of a cluster.
  • Cluster: A set of Nodes that run containerized applications managed by Kubernetes. For this example, and in most common Kubernetes deployments, nodes in the cluster are not part of the public internet.
  • Edge router: A router that enforces the firewall policy for your cluster. This could be a gateway managed by a cloud provider or a physical piece of hardware.
  • Cluster network: A set of links, logical or physical, that facilitate communication within a cluster according to the Kubernetes networking model.
  • Service: A Kubernetes Service that identifies a set of Pods using label selectors. Unless mentioned otherwise, Services are assumed to have virtual IPs only routable within the cluster network.

What is Ingress?

Ingress exposes HTTP and HTTPS routes from outside the cluster to services within the cluster. Traffic routing is controlled by rules defined on the Ingress resource.

Here is a simple example where an Ingress sends all its traffic to one Service:

ingress-diagram

Figure. Ingress

An Ingress may be configured to give Services externally-reachable URLs, load balance traffic, terminate SSL / TLS, and offer name-based virtual hosting. An Ingress controller is responsible for fulfilling the Ingress, usually with a load balancer, though it may also configure your edge router or additional frontends to help handle the traffic.

An Ingress does not expose arbitrary ports or protocols. Exposing services other than HTTP and HTTPS to the internet typically uses a service of type Service.Type=NodePort or Service.Type=LoadBalancer.

Prerequisites

You must have an Ingress controller to satisfy an Ingress. Only creating an Ingress resource has no effect.

You may need to deploy an Ingress controller such as ingress-nginx. You can choose from a number of Ingress controllers.

Ideally, all Ingress controllers should fit the reference specification. In reality, the various Ingress controllers operate slightly differently.

The Ingress resource

A minimal Ingress resource example:

apiVersion: networking.k8s.io/v1
kind: Ingress
metadata:
  name: minimal-ingress
  annotations:
    nginx.ingress.kubernetes.io/rewrite-target: /
spec:
  ingressClassName: nginx-example
  rules:
  - http:
      paths:
      - path: /testpath
        pathType: Prefix
        backend:
          service:
            name: test
            port:
              number: 80

An Ingress needs apiVersion, kind, metadata and spec fields. The name of an Ingress object must be a valid DNS subdomain name. For general information about working with config files, see deploying applications, configuring containers, managing resources. Ingress frequently uses annotations to configure some options depending on the Ingress controller, an example of which is the rewrite-target annotation. Different Ingress controllers support different annotations. Review the documentation for your choice of Ingress controller to learn which annotations are supported.

The Ingress spec has all the information needed to configure a load balancer or proxy server. Most importantly, it contains a list of rules matched against all incoming requests. Ingress resource only supports rules for directing HTTP(S) traffic.

If the ingressClassName is omitted, a default Ingress class should be defined.

There are some ingress controllers, that work without the definition of a default IngressClass. For example, the Ingress-NGINX controller can be configured with a flag --watch-ingress-without-class. It is recommended though, to specify the default IngressClass as shown below.

Ingress rules

Each HTTP rule contains the following information:

  • An optional host. In this example, no host is specified, so the rule applies to all inbound HTTP traffic through the IP address specified. If a host is provided (for example, foo.bar.com), the rules apply to that host.
  • A list of paths (for example, /testpath), each of which has an associated backend defined with a service.name and a service.port.name or service.port.number. Both the host and path must match the content of an incoming request before the load balancer directs traffic to the referenced Service.
  • A backend is a combination of Service and port names as described in the Service doc or a custom resource backend by way of a CRD. HTTP (and HTTPS) requests to the Ingress that matches the host and path of the rule are sent to the listed backend.

A defaultBackend is often configured in an Ingress controller to service any requests that do not match a path in the spec.

DefaultBackend

An Ingress with no rules sends all traffic to a single default backend and .spec.defaultBackend is the backend that should handle requests in that case. The defaultBackend is conventionally a configuration option of the Ingress controller and is not specified in your Ingress resources. If no .spec.rules are specified, .spec.defaultBackend must be specified. If defaultBackend is not set, the handling of requests that do not match any of the rules will be up to the ingress controller (consult the documentation for your ingress controller to find out how it handles this case).

If none of the hosts or paths match the HTTP request in the Ingress objects, the traffic is routed to your default backend.

Resource backends

A Resource backend is an ObjectRef to another Kubernetes resource within the same namespace as the Ingress object. A Resource is a mutually exclusive setting with Service, and will fail validation if both are specified. A common usage for a Resource backend is to ingress data to an object storage backend with static assets.

apiVersion: networking.k8s.io/v1
kind: Ingress
metadata:
  name: ingress-resource-backend
spec:
  defaultBackend:
    resource:
      apiGroup: k8s.example.com
      kind: StorageBucket
      name: static-assets
  rules:
    - http:
        paths:
          - path: /icons
            pathType: ImplementationSpecific
            backend:
              resource:
                apiGroup: k8s.example.com
                kind: StorageBucket
                name: icon-assets

After creating the Ingress above, you can view it with the following command:

kubectl describe ingress ingress-resource-backend
Name:             ingress-resource-backend
Namespace:        default
Address:
Default backend:  APIGroup: k8s.example.com, Kind: StorageBucket, Name: static-assets
Rules:
  Host        Path  Backends
  ----        ----  --------
  *
              /icons   APIGroup: k8s.example.com, Kind: StorageBucket, Name: icon-assets
Annotations:  <none>
Events:       <none>

Path types

Each path in an Ingress is required to have a corresponding path type. Paths that do not include an explicit pathType will fail validation. There are three supported path types:

  • ImplementationSpecific: With this path type, matching is up to the IngressClass. Implementations can treat this as a separate pathType or treat it identically to Prefix or Exact path types.

  • Exact: Matches the URL path exactly and with case sensitivity.

  • Prefix: Matches based on a URL path prefix split by /. Matching is case sensitive and done on a path element by element basis. A path element refers to the list of labels in the path split by the / separator. A request is a match for path p if every p is an element-wise prefix of p of the request path.

Examples

Kind Path(s) Request path(s) Matches?
Prefix / (all paths) Yes
Exact /foo /foo Yes
Exact /foo /bar No
Exact /foo /foo/ No
Exact /foo/ /foo No
Prefix /foo /foo, /foo/ Yes
Prefix /foo/ /foo, /foo/ Yes
Prefix /aaa/bb /aaa/bbb No
Prefix /aaa/bbb /aaa/bbb Yes
Prefix /aaa/bbb/ /aaa/bbb Yes, ignores trailing slash
Prefix /aaa/bbb /aaa/bbb/ Yes, matches trailing slash
Prefix /aaa/bbb /aaa/bbb/ccc Yes, matches subpath
Prefix /aaa/bbb /aaa/bbbxyz No, does not match string prefix
Prefix /, /aaa /aaa/ccc Yes, matches /aaa prefix
Prefix /, /aaa, /aaa/bbb /aaa/bbb Yes, matches /aaa/bbb prefix
Prefix /, /aaa, /aaa/bbb /ccc Yes, matches / prefix
Prefix /aaa /ccc No, uses default backend
Mixed /foo (Prefix), /foo (Exact) /foo Yes, prefers Exact

Multiple matches

In some cases, multiple paths within an Ingress will match a request. In those cases precedence will be given first to the longest matching path. If two paths are still equally matched, precedence will be given to paths with an exact path type over prefix path type.

Hostname wildcards

Hosts can be precise matches (for example “foo.bar.com”) or a wildcard (for example “*.foo.com”). Precise matches require that the HTTP host header matches the host field. Wildcard matches require the HTTP host header is equal to the suffix of the wildcard rule.

Host Host header Match?
*.foo.com bar.foo.com Matches based on shared suffix
*.foo.com baz.bar.foo.com No match, wildcard only covers a single DNS label
*.foo.com foo.com No match, wildcard only covers a single DNS label
apiVersion: networking.k8s.io/v1
kind: Ingress
metadata:
  name: ingress-wildcard-host
spec:
  rules:
  - host: "foo.bar.com"
    http:
      paths:
      - pathType: Prefix
        path: "/bar"
        backend:
          service:
            name: service1
            port:
              number: 80
  - host: "*.foo.com"
    http:
      paths:
      - pathType: Prefix
        path: "/foo"
        backend:
          service:
            name: service2
            port:
              number: 80

Ingress class

Ingresses can be implemented by different controllers, often with different configuration. Each Ingress should specify a class, a reference to an IngressClass resource that contains additional configuration including the name of the controller that should implement the class.

apiVersion: networking.k8s.io/v1
kind: IngressClass
metadata:
  name: external-lb
spec:
  controller: example.com/ingress-controller
  parameters:
    apiGroup: k8s.example.com
    kind: IngressParameters
    name: external-lb

The .spec.parameters field of an IngressClass lets you reference another resource that provides configuration related to that IngressClass.

The specific type of parameters to use depends on the ingress controller that you specify in the .spec.controller field of the IngressClass.

IngressClass scope

Depending on your ingress controller, you may be able to use parameters that you set cluster-wide, or just for one namespace.

The default scope for IngressClass parameters is cluster-wide.

If you set the .spec.parameters field and don't set .spec.parameters.scope, or if you set .spec.parameters.scope to Cluster, then the IngressClass refers to a cluster-scoped resource. The kind (in combination the apiGroup) of the parameters refers to a cluster-scoped API (possibly a custom resource), and the name of the parameters identifies a specific cluster scoped resource for that API.

For example:

---
apiVersion: networking.k8s.io/v1
kind: IngressClass
metadata:
  name: external-lb-1
spec:
  controller: example.com/ingress-controller
  parameters:
    # The parameters for this IngressClass are specified in a
    # ClusterIngressParameter (API group k8s.example.net) named
    # "external-config-1". This definition tells Kubernetes to
    # look for a cluster-scoped parameter resource.
    scope: Cluster
    apiGroup: k8s.example.net
    kind: ClusterIngressParameter
    name: external-config-1

FEATURE STATE: Kubernetes v1.23 [stable]

If you set the .spec.parameters field and set .spec.parameters.scope to Namespace, then the IngressClass refers to a namespaced-scoped resource. You must also set the namespace field within .spec.parameters to the namespace that contains the parameters you want to use.

The kind (in combination the apiGroup) of the parameters refers to a namespaced API (for example: ConfigMap), and the name of the parameters identifies a specific resource in the namespace you specified in namespace.

Namespace-scoped parameters help the cluster operator delegate control over the configuration (for example: load balancer settings, API gateway definition) that is used for a workload. If you used a cluster-scoped parameter then either:

  • the cluster operator team needs to approve a different team's changes every time there's a new configuration change being applied.
  • the cluster operator must define specific access controls, such as RBAC roles and bindings, that let the application team make changes to the cluster-scoped parameters resource.

The IngressClass API itself is always cluster-scoped.

Here is an example of an IngressClass that refers to parameters that are namespaced:

---
apiVersion: networking.k8s.io/v1
kind: IngressClass
metadata:
  name: external-lb-2
spec:
  controller: example.com/ingress-controller
  parameters:
    # The parameters for this IngressClass are specified in an
    # IngressParameter (API group k8s.example.com) named "external-config",
    # that's in the "external-configuration" namespace.
    scope: Namespace
    apiGroup: k8s.example.com
    kind: IngressParameter
    namespace: external-configuration
    name: external-config

Deprecated annotation

Before the IngressClass resource and ingressClassName field were added in Kubernetes 1.18, Ingress classes were specified with a kubernetes.io/ingress.class annotation on the Ingress. This annotation was never formally defined, but was widely supported by Ingress controllers.

The newer ingressClassName field on Ingresses is a replacement for that annotation, but is not a direct equivalent. While the annotation was generally used to reference the name of the Ingress controller that should implement the Ingress, the field is a reference to an IngressClass resource that contains additional Ingress configuration, including the name of the Ingress controller.

Default IngressClass

You can mark a particular IngressClass as default for your cluster. Setting the ingressclass.kubernetes.io/is-default-class annotation to true on an IngressClass resource will ensure that new Ingresses without an ingressClassName field specified will be assigned this default IngressClass.

There are some ingress controllers, that work without the definition of a default IngressClass. For example, the Ingress-NGINX controller can be configured with a flag --watch-ingress-without-class. It is recommended though, to specify the default IngressClass:

apiVersion: networking.k8s.io/v1
kind: IngressClass
metadata:
  labels:
    app.kubernetes.io/component: controller
  name: nginx-example
  annotations:
    ingressclass.kubernetes.io/is-default-class: "true"
spec:
  controller: k8s.io/ingress-nginx

Types of Ingress

Ingress backed by a single Service

There are existing Kubernetes concepts that allow you to expose a single Service (see alternatives). You can also do this with an Ingress by specifying a default backend with no rules.

apiVersion: networking.k8s.io/v1
kind: Ingress
metadata:
  name: test-ingress
spec:
  defaultBackend:
    service:
      name: test
      port:
        number: 80

If you create it using kubectl apply -f you should be able to view the state of the Ingress you added:

kubectl get ingress test-ingress
NAME           CLASS         HOSTS   ADDRESS         PORTS   AGE
test-ingress   external-lb   *       203.0.113.123   80      59s

Where 203.0.113.123 is the IP allocated by the Ingress controller to satisfy this Ingress.

Simple fanout

A fanout configuration routes traffic from a single IP address to more than one Service, based on the HTTP URI being requested. An Ingress allows you to keep the number of load balancers down to a minimum. For example, a setup like:

ingress-fanout-diagram

Figure. Ingress Fan Out

would require an Ingress such as:

apiVersion: networking.k8s.io/v1
kind: Ingress
metadata:
  name: simple-fanout-example
spec:
  rules:
  - host: foo.bar.com
    http:
      paths:
      - path: /foo
        pathType: Prefix
        backend:
          service:
            name: service1
            port:
              number: 4200
      - path: /bar
        pathType: Prefix
        backend:
          service:
            name: service2
            port:
              number: 8080

When you create the Ingress with kubectl apply -f:

kubectl describe ingress simple-fanout-example
Name:             simple-fanout-example
Namespace:        default
Address:          178.91.123.132
Default backend:  default-http-backend:80 (10.8.2.3:8080)
Rules:
  Host         Path  Backends
  ----         ----  --------
  foo.bar.com
               /foo   service1:4200 (10.8.0.90:4200)
               /bar   service2:8080 (10.8.0.91:8080)
Events:
  Type     Reason  Age                From                     Message
  ----     ------  ----               ----                     -------
  Normal   ADD     22s                loadbalancer-controller  default/test

The Ingress controller provisions an implementation-specific load balancer that satisfies the Ingress, as long as the Services (service1, service2) exist. When it has done so, you can see the address of the load balancer at the Address field.

Name based virtual hosting

Name-based virtual hosts support routing HTTP traffic to multiple host names at the same IP address.

ingress-namebase-diagram

Figure. Ingress Name Based Virtual hosting

The following Ingress tells the backing load balancer to route requests based on the Host header.

apiVersion: networking.k8s.io/v1
kind: Ingress
metadata:
  name: name-virtual-host-ingress
spec:
  rules:
  - host: foo.bar.com
    http:
      paths:
      - pathType: Prefix
        path: "/"
        backend:
          service:
            name: service1
            port:
              number: 80
  - host: bar.foo.com
    http:
      paths:
      - pathType: Prefix
        path: "/"
        backend:
          service:
            name: service2
            port:
              number: 80

If you create an Ingress resource without any hosts defined in the rules, then any web traffic to the IP address of your Ingress controller can be matched without a name based virtual host being required.

For example, the following Ingress routes traffic requested for first.bar.com to service1, second.bar.com to service2, and any traffic whose request host header doesn't match first.bar.com and second.bar.com to service3.

apiVersion: networking.k8s.io/v1
kind: Ingress
metadata:
  name: name-virtual-host-ingress-no-third-host
spec:
  rules:
  - host: first.bar.com
    http:
      paths:
      - pathType: Prefix
        path: "/"
        backend:
          service:
            name: service1
            port:
              number: 80
  - host: second.bar.com
    http:
      paths:
      - pathType: Prefix
        path: "/"
        backend:
          service:
            name: service2
            port:
              number: 80
  - http:
      paths:
      - pathType: Prefix
        path: "/"
        backend:
          service:
            name: service3
            port:
              number: 80

TLS

You can secure an Ingress by specifying a Secret that contains a TLS private key and certificate. The Ingress resource only supports a single TLS port, 443, and assumes TLS termination at the ingress point (traffic to the Service and its Pods is in plaintext). If the TLS configuration section in an Ingress specifies different hosts, they are multiplexed on the same port according to the hostname specified through the SNI TLS extension (provided the Ingress controller supports SNI). The TLS secret must contain keys named tls.crt and tls.key that contain the certificate and private key to use for TLS. For example:

apiVersion: v1
kind: Secret
metadata:
  name: testsecret-tls
  namespace: default
data:
  tls.crt: base64 encoded cert
  tls.key: base64 encoded key
type: kubernetes.io/tls

Referencing this secret in an Ingress tells the Ingress controller to secure the channel from the client to the load balancer using TLS. You need to make sure the TLS secret you created came from a certificate that contains a Common Name (CN), also known as a Fully Qualified Domain Name (FQDN) for https-example.foo.com.

apiVersion: networking.k8s.io/v1
kind: Ingress
metadata:
  name: tls-example-ingress
spec:
  tls:
  - hosts:
      - https-example.foo.com
    secretName: testsecret-tls
  rules:
  - host: https-example.foo.com
    http:
      paths:
      - path: /
        pathType: Prefix
        backend:
          service:
            name: service1
            port:
              number: 80

Load balancing

An Ingress controller is bootstrapped with some load balancing policy settings that it applies to all Ingress, such as the load balancing algorithm, backend weight scheme, and others. More advanced load balancing concepts (e.g. persistent sessions, dynamic weights) are not yet exposed through the Ingress. You can instead get these features through the load balancer used for a Service.

It's also worth noting that even though health checks are not exposed directly through the Ingress, there exist parallel concepts in Kubernetes such as readiness probes that allow you to achieve the same end result. Please review the controller specific documentation to see how they handle health checks (for example: nginx, or GCE).

Updating an Ingress

To update an existing Ingress to add a new Host, you can update it by editing the resource:

kubectl describe ingress test
Name:             test
Namespace:        default
Address:          178.91.123.132
Default backend:  default-http-backend:80 (10.8.2.3:8080)
Rules:
  Host         Path  Backends
  ----         ----  --------
  foo.bar.com
               /foo   service1:80 (10.8.0.90:80)
Annotations:
  nginx.ingress.kubernetes.io/rewrite-target:  /
Events:
  Type     Reason  Age                From                     Message
  ----     ------  ----               ----                     -------
  Normal   ADD     35s                loadbalancer-controller  default/test
kubectl edit ingress test

This pops up an editor with the existing configuration in YAML format. Modify it to include the new Host:

spec:
  rules:
  - host: foo.bar.com
    http:
      paths:
      - backend:
          service:
            name: service1
            port:
              number: 80
        path: /foo
        pathType: Prefix
  - host: bar.baz.com
    http:
      paths:
      - backend:
          service:
            name: service2
            port:
              number: 80
        path: /foo
        pathType: Prefix
..

After you save your changes, kubectl updates the resource in the API server, which tells the Ingress controller to reconfigure the load balancer.

Verify this:

kubectl describe ingress test
Name:             test
Namespace:        default
Address:          178.91.123.132
Default backend:  default-http-backend:80 (10.8.2.3:8080)
Rules:
  Host         Path  Backends
  ----         ----  --------
  foo.bar.com
               /foo   service1:80 (10.8.0.90:80)
  bar.baz.com
               /foo   service2:80 (10.8.0.91:80)
Annotations:
  nginx.ingress.kubernetes.io/rewrite-target:  /
Events:
  Type     Reason  Age                From                     Message
  ----     ------  ----               ----                     -------
  Normal   ADD     45s                loadbalancer-controller  default/test

You can achieve the same outcome by invoking kubectl replace -f on a modified Ingress YAML file.

Failing across availability zones

Techniques for spreading traffic across failure domains differ between cloud providers. Please check the documentation of the relevant Ingress controller for details.

Alternatives

You can expose a Service in multiple ways that don't directly involve the Ingress resource:

What's next

3 - Ingress Controllers

In order for an Ingress to work in your cluster, there must be an ingress controller running. You need to select at least one ingress controller and make sure it is set up in your cluster. This page lists common ingress controllers that you can deploy.

In order for the Ingress resource to work, the cluster must have an ingress controller running.

Unlike other types of controllers which run as part of the kube-controller-manager binary, Ingress controllers are not started automatically with a cluster. Use this page to choose the ingress controller implementation that best fits your cluster.

Kubernetes as a project supports and maintains AWS, GCE, and nginx ingress controllers.

Additional controllers

Using multiple Ingress controllers

You may deploy any number of ingress controllers using ingress class within a cluster. Note the .metadata.name of your ingress class resource. When you create an ingress you would need that name to specify the ingressClassName field on your Ingress object (refer to IngressSpec v1 reference. ingressClassName is a replacement of the older annotation method.

If you do not specify an IngressClass for an Ingress, and your cluster has exactly one IngressClass marked as default, then Kubernetes applies the cluster's default IngressClass to the Ingress. You mark an IngressClass as default by setting the ingressclass.kubernetes.io/is-default-class annotation on that IngressClass, with the string value "true".

Ideally, all ingress controllers should fulfill this specification, but the various ingress controllers operate slightly differently.

What's next

4 - EndpointSlices

The EndpointSlice API is the mechanism that Kubernetes uses to let your Service scale to handle large numbers of backends, and allows the cluster to update its list of healthy backends efficiently.
FEATURE STATE: Kubernetes v1.21 [stable]

Kubernetes' EndpointSlice API provides a way to track network endpoints within a Kubernetes cluster. EndpointSlices offer a more scalable and extensible alternative to Endpoints.

EndpointSlice API

In Kubernetes, an EndpointSlice contains references to a set of network endpoints. The control plane automatically creates EndpointSlices for any Kubernetes Service that has a selector specified. These EndpointSlices include references to all the Pods that match the Service selector. EndpointSlices group network endpoints together by unique combinations of protocol, port number, and Service name. The name of a EndpointSlice object must be a valid DNS subdomain name.

As an example, here's a sample EndpointSlice object, that's owned by the example Kubernetes Service.

apiVersion: discovery.k8s.io/v1
kind: EndpointSlice
metadata:
  name: example-abc
  labels:
    kubernetes.io/service-name: example
addressType: IPv4
ports:
  - name: http
    protocol: TCP
    port: 80
endpoints:
  - addresses:
      - "10.1.2.3"
    conditions:
      ready: true
    hostname: pod-1
    nodeName: node-1
    zone: us-west2-a

By default, the control plane creates and manages EndpointSlices to have no more than 100 endpoints each. You can configure this with the --max-endpoints-per-slice kube-controller-manager flag, up to a maximum of 1000.

EndpointSlices can act as the source of truth for kube-proxy when it comes to how to route internal traffic.

Address types

EndpointSlices support three address types:

  • IPv4
  • IPv6
  • FQDN (Fully Qualified Domain Name)

Each EndpointSlice object represents a specific IP address type. If you have a Service that is available via IPv4 and IPv6, there will be at least two EndpointSlice objects (one for IPv4, and one for IPv6).

Conditions

The EndpointSlice API stores conditions about endpoints that may be useful for consumers. The three conditions are ready, serving, and terminating.

Ready

ready is a condition that maps to a Pod's Ready condition. A running Pod with the Ready condition set to True should have this EndpointSlice condition also set to true. For compatibility reasons, ready is NEVER true when a Pod is terminating. Consumers should refer to the serving condition to inspect the readiness of terminating Pods. The only exception to this rule is for Services with spec.publishNotReadyAddresses set to true. Endpoints for these Services will always have the ready condition set to true.

Serving

FEATURE STATE: Kubernetes v1.22 [beta]

serving is identical to the ready condition, except it does not account for terminating states. Consumers of the EndpointSlice API should check this condition if they care about pod readiness while the pod is also terminating.

Terminating

FEATURE STATE: Kubernetes v1.22 [beta]

Terminating is a condition that indicates whether an endpoint is terminating. For pods, this is any pod that has a deletion timestamp set.

Topology information

Each endpoint within an EndpointSlice can contain relevant topology information. The topology information includes the location of the endpoint and information about the corresponding Node and zone. These are available in the following per endpoint fields on EndpointSlices:

  • nodeName - The name of the Node this endpoint is on.
  • zone - The zone this endpoint is in.

Management

Most often, the control plane (specifically, the endpoint slice controller) creates and manages EndpointSlice objects. There are a variety of other use cases for EndpointSlices, such as service mesh implementations, that could result in other entities or controllers managing additional sets of EndpointSlices.

To ensure that multiple entities can manage EndpointSlices without interfering with each other, Kubernetes defines the label endpointslice.kubernetes.io/managed-by, which indicates the entity managing an EndpointSlice. The endpoint slice controller sets endpointslice-controller.k8s.io as the value for this label on all EndpointSlices it manages. Other entities managing EndpointSlices should also set a unique value for this label.

Ownership

In most use cases, EndpointSlices are owned by the Service that the endpoint slice object tracks endpoints for. This ownership is indicated by an owner reference on each EndpointSlice as well as a kubernetes.io/service-name label that enables simple lookups of all EndpointSlices belonging to a Service.

EndpointSlice mirroring

In some cases, applications create custom Endpoints resources. To ensure that these applications do not need to concurrently write to both Endpoints and EndpointSlice resources, the cluster's control plane mirrors most Endpoints resources to corresponding EndpointSlices.

The control plane mirrors Endpoints resources unless:

  • the Endpoints resource has a endpointslice.kubernetes.io/skip-mirror label set to true.
  • the Endpoints resource has a control-plane.alpha.kubernetes.io/leader annotation.
  • the corresponding Service resource does not exist.
  • the corresponding Service resource has a non-nil selector.

Individual Endpoints resources may translate into multiple EndpointSlices. This will occur if an Endpoints resource has multiple subsets or includes endpoints with multiple IP families (IPv4 and IPv6). A maximum of 1000 addresses per subset will be mirrored to EndpointSlices.

Distribution of EndpointSlices

Each EndpointSlice has a set of ports that applies to all endpoints within the resource. When named ports are used for a Service, Pods may end up with different target port numbers for the same named port, requiring different EndpointSlices. This is similar to the logic behind how subsets are grouped with Endpoints.

The control plane tries to fill EndpointSlices as full as possible, but does not actively rebalance them. The logic is fairly straightforward:

  1. Iterate through existing EndpointSlices, remove endpoints that are no longer desired and update matching endpoints that have changed.
  2. Iterate through EndpointSlices that have been modified in the first step and fill them up with any new endpoints needed.
  3. If there's still new endpoints left to add, try to fit them into a previously unchanged slice and/or create new ones.

Importantly, the third step prioritizes limiting EndpointSlice updates over a perfectly full distribution of EndpointSlices. As an example, if there are 10 new endpoints to add and 2 EndpointSlices with room for 5 more endpoints each, this approach will create a new EndpointSlice instead of filling up the 2 existing EndpointSlices. In other words, a single EndpointSlice creation is preferrable to multiple EndpointSlice updates.

With kube-proxy running on each Node and watching EndpointSlices, every change to an EndpointSlice becomes relatively expensive since it will be transmitted to every Node in the cluster. This approach is intended to limit the number of changes that need to be sent to every Node, even if it may result with multiple EndpointSlices that are not full.

In practice, this less than ideal distribution should be rare. Most changes processed by the EndpointSlice controller will be small enough to fit in an existing EndpointSlice, and if not, a new EndpointSlice is likely going to be necessary soon anyway. Rolling updates of Deployments also provide a natural repacking of EndpointSlices with all Pods and their corresponding endpoints getting replaced.

Duplicate endpoints

Due to the nature of EndpointSlice changes, endpoints may be represented in more than one EndpointSlice at the same time. This naturally occurs as changes to different EndpointSlice objects can arrive at the Kubernetes client watch / cache at different times.

Comparison with Endpoints

The original Endpoints API provided a simple and straightforward way of tracking network endpoints in Kubernetes. As Kubernetes clusters and Services grew to handle more traffic and to send more traffic to more backend Pods, the limitations of that original API became more visible. Most notably, those included challenges with scaling to larger numbers of network endpoints.

Since all network endpoints for a Service were stored in a single Endpoints object, those Endpoints objects could get quite large. For Services that stayed stable (the same set of endpoints over a long period of time) the impact was less noticeable; even then, some use cases of Kubernetes weren't well served.

When a Service had a lot of backend endpoints and the workload was either scaling frequently, or rolling out new changes frequently, each update to the single Endpoints object for that Service meant a lot of traffic between Kubernetes cluster components (within the control plane, and also between nodes and the API server). This extra traffic also had a cost in terms of CPU use.

With EndpointSlices, adding or removing a single Pod triggers the same number of updates to clients that are watching for changes, but the size of those update message is much smaller at large scale.

EndpointSlices also enabled innovation around new features such dual-stack networking and topology-aware routing.

What's next

5 - Network Policies

If you want to control traffic flow at the IP address or port level (OSI layer 3 or 4), NetworkPolicies allow you to specify rules for traffic flow within your cluster, and also between Pods and the outside world. Your cluster must use a network plugin that supports NetworkPolicy enforcement.

If you want to control traffic flow at the IP address or port level (OSI layer 3 or 4), then you might consider using Kubernetes NetworkPolicies for particular applications in your cluster. NetworkPolicies are an application-centric construct which allow you to specify how a pod is allowed to communicate with various network "entities" (we use the word "entity" here to avoid overloading the more common terms such as "endpoints" and "services", which have specific Kubernetes connotations) over the network. NetworkPolicies apply to a connection with a pod on one or both ends, and are not relevant to other connections.

The entities that a Pod can communicate with are identified through a combination of the following 3 identifiers:

  1. Other pods that are allowed (exception: a pod cannot block access to itself)
  2. Namespaces that are allowed
  3. IP blocks (exception: traffic to and from the node where a Pod is running is always allowed, regardless of the IP address of the Pod or the node)

When defining a pod- or namespace- based NetworkPolicy, you use a selector to specify what traffic is allowed to and from the Pod(s) that match the selector.

Meanwhile, when IP based NetworkPolicies are created, we define policies based on IP blocks (CIDR ranges).

Prerequisites

Network policies are implemented by the network plugin. To use network policies, you must be using a networking solution which supports NetworkPolicy. Creating a NetworkPolicy resource without a controller that implements it will have no effect.

The Two Sorts of Pod Isolation

There are two sorts of isolation for a pod: isolation for egress, and isolation for ingress. They concern what connections may be established. "Isolation" here is not absolute, rather it means "some restrictions apply". The alternative, "non-isolated for $direction", means that no restrictions apply in the stated direction. The two sorts of isolation (or not) are declared independently, and are both relevant for a connection from one pod to another.

By default, a pod is non-isolated for egress; all outbound connections are allowed. A pod is isolated for egress if there is any NetworkPolicy that both selects the pod and has "Egress" in its policyTypes; we say that such a policy applies to the pod for egress. When a pod is isolated for egress, the only allowed connections from the pod are those allowed by the egress list of some NetworkPolicy that applies to the pod for egress. The effects of those egress lists combine additively.

By default, a pod is non-isolated for ingress; all inbound connections are allowed. A pod is isolated for ingress if there is any NetworkPolicy that both selects the pod and has "Ingress" in its policyTypes; we say that such a policy applies to the pod for ingress. When a pod is isolated for ingress, the only allowed connections into the pod are those from the pod's node and those allowed by the ingress list of some NetworkPolicy that applies to the pod for ingress. The effects of those ingress lists combine additively.

Network policies do not conflict; they are additive. If any policy or policies apply to a given pod for a given direction, the connections allowed in that direction from that pod is the union of what the applicable policies allow. Thus, order of evaluation does not affect the policy result.

For a connection from a source pod to a destination pod to be allowed, both the egress policy on the source pod and the ingress policy on the destination pod need to allow the connection. If either side does not allow the connection, it will not happen.

The NetworkPolicy resource

See the NetworkPolicy reference for a full definition of the resource.

An example NetworkPolicy might look like this:

apiVersion: networking.k8s.io/v1
kind: NetworkPolicy
metadata:
  name: test-network-policy
  namespace: default
spec:
  podSelector:
    matchLabels:
      role: db
  policyTypes:
    - Ingress
    - Egress
  ingress:
    - from:
        - ipBlock:
            cidr: 172.17.0.0/16
            except:
              - 172.17.1.0/24
        - namespaceSelector:
            matchLabels:
              project: myproject
        - podSelector:
            matchLabels:
              role: frontend
      ports:
        - protocol: TCP
          port: 6379
  egress:
    - to:
        - ipBlock:
            cidr: 10.0.0.0/24
      ports:
        - protocol: TCP
          port: 5978

Mandatory Fields: As with all other Kubernetes config, a NetworkPolicy needs apiVersion, kind, and metadata fields. For general information about working with config files, see Configure a Pod to Use a ConfigMap, and Object Management.

spec: NetworkPolicy spec has all the information needed to define a particular network policy in the given namespace.

podSelector: Each NetworkPolicy includes a podSelector which selects the grouping of pods to which the policy applies. The example policy selects pods with the label "role=db". An empty podSelector selects all pods in the namespace.

policyTypes: Each NetworkPolicy includes a policyTypes list which may include either Ingress, Egress, or both. The policyTypes field indicates whether or not the given policy applies to ingress traffic to selected pod, egress traffic from selected pods, or both. If no policyTypes are specified on a NetworkPolicy then by default Ingress will always be set and Egress will be set if the NetworkPolicy has any egress rules.

ingress: Each NetworkPolicy may include a list of allowed ingress rules. Each rule allows traffic which matches both the from and ports sections. The example policy contains a single rule, which matches traffic on a single port, from one of three sources, the first specified via an ipBlock, the second via a namespaceSelector and the third via a podSelector.

egress: Each NetworkPolicy may include a list of allowed egress rules. Each rule allows traffic which matches both the to and ports sections. The example policy contains a single rule, which matches traffic on a single port to any destination in 10.0.0.0/24.

So, the example NetworkPolicy:

  1. isolates "role=db" pods in the "default" namespace for both ingress and egress traffic (if they weren't already isolated)

  2. (Ingress rules) allows connections to all pods in the "default" namespace with the label "role=db" on TCP port 6379 from:

    • any pod in the "default" namespace with the label "role=frontend"
    • any pod in a namespace with the label "project=myproject"
    • IP addresses in the ranges 172.17.0.0–172.17.0.255 and 172.17.2.0–172.17.255.255 (ie, all of 172.17.0.0/16 except 172.17.1.0/24)
  3. (Egress rules) allows connections from any pod in the "default" namespace with the label "role=db" to CIDR 10.0.0.0/24 on TCP port 5978

See the Declare Network Policy walkthrough for further examples.

Behavior of to and from selectors

There are four kinds of selectors that can be specified in an ingress from section or egress to section:

podSelector: This selects particular Pods in the same namespace as the NetworkPolicy which should be allowed as ingress sources or egress destinations.

namespaceSelector: This selects particular namespaces for which all Pods should be allowed as ingress sources or egress destinations.

namespaceSelector and podSelector: A single to/from entry that specifies both namespaceSelector and podSelector selects particular Pods within particular namespaces. Be careful to use correct YAML syntax; this policy:

  ...
  ingress:
  - from:
    - namespaceSelector:
        matchLabels:
          user: alice
      podSelector:
        matchLabels:
          role: client
  ...

contains a single from element allowing connections from Pods with the label role=client in namespaces with the label user=alice. But this policy:

  ...
  ingress:
  - from:
    - namespaceSelector:
        matchLabels:
          user: alice
    - podSelector:
        matchLabels:
          role: client
  ...

contains two elements in the from array, and allows connections from Pods in the local Namespace with the label role=client, or from any Pod in any namespace with the label user=alice.

When in doubt, use kubectl describe to see how Kubernetes has interpreted the policy.

ipBlock: This selects particular IP CIDR ranges to allow as ingress sources or egress destinations. These should be cluster-external IPs, since Pod IPs are ephemeral and unpredictable.

Cluster ingress and egress mechanisms often require rewriting the source or destination IP of packets. In cases where this happens, it is not defined whether this happens before or after NetworkPolicy processing, and the behavior may be different for different combinations of network plugin, cloud provider, Service implementation, etc.

In the case of ingress, this means that in some cases you may be able to filter incoming packets based on the actual original source IP, while in other cases, the "source IP" that the NetworkPolicy acts on may be the IP of a LoadBalancer or of the Pod's node, etc.

For egress, this means that connections from pods to Service IPs that get rewritten to cluster-external IPs may or may not be subject to ipBlock-based policies.

Default policies

By default, if no policies exist in a namespace, then all ingress and egress traffic is allowed to and from pods in that namespace. The following examples let you change the default behavior in that namespace.

Default deny all ingress traffic

You can create a "default" ingress isolation policy for a namespace by creating a NetworkPolicy that selects all pods but does not allow any ingress traffic to those pods.

---
apiVersion: networking.k8s.io/v1
kind: NetworkPolicy
metadata:
  name: default-deny-ingress
spec:
  podSelector: {}
  policyTypes:
  - Ingress

This ensures that even pods that aren't selected by any other NetworkPolicy will still be isolated for ingress. This policy does not affect isolation for egress from any pod.

Allow all ingress traffic

If you want to allow all incoming connections to all pods in a namespace, you can create a policy that explicitly allows that.

---
apiVersion: networking.k8s.io/v1
kind: NetworkPolicy
metadata:
  name: allow-all-ingress
spec:
  podSelector: {}
  ingress:
  - {}
  policyTypes:
  - Ingress

With this policy in place, no additional policy or policies can cause any incoming connection to those pods to be denied. This policy has no effect on isolation for egress from any pod.

Default deny all egress traffic

You can create a "default" egress isolation policy for a namespace by creating a NetworkPolicy that selects all pods but does not allow any egress traffic from those pods.

---
apiVersion: networking.k8s.io/v1
kind: NetworkPolicy
metadata:
  name: default-deny-egress
spec:
  podSelector: {}
  policyTypes:
  - Egress

This ensures that even pods that aren't selected by any other NetworkPolicy will not be allowed egress traffic. This policy does not change the ingress isolation behavior of any pod.

Allow all egress traffic

If you want to allow all connections from all pods in a namespace, you can create a policy that explicitly allows all outgoing connections from pods in that namespace.

---
apiVersion: networking.k8s.io/v1
kind: NetworkPolicy
metadata:
  name: allow-all-egress
spec:
  podSelector: {}
  egress:
  - {}
  policyTypes:
  - Egress

With this policy in place, no additional policy or policies can cause any outgoing connection from those pods to be denied. This policy has no effect on isolation for ingress to any pod.

Default deny all ingress and all egress traffic

You can create a "default" policy for a namespace which prevents all ingress AND egress traffic by creating the following NetworkPolicy in that namespace.

---
apiVersion: networking.k8s.io/v1
kind: NetworkPolicy
metadata:
  name: default-deny-all
spec:
  podSelector: {}
  policyTypes:
  - Ingress
  - Egress

This ensures that even pods that aren't selected by any other NetworkPolicy will not be allowed ingress or egress traffic.

SCTP support

FEATURE STATE: Kubernetes v1.20 [stable]

As a stable feature, this is enabled by default. To disable SCTP at a cluster level, you (or your cluster administrator) will need to disable the SCTPSupport feature gate for the API server with --feature-gates=SCTPSupport=false,…. When the feature gate is enabled, you can set the protocol field of a NetworkPolicy to SCTP.

Targeting a range of ports

FEATURE STATE: Kubernetes v1.25 [stable]

When writing a NetworkPolicy, you can target a range of ports instead of a single port.

This is achievable with the usage of the endPort field, as the following example:

apiVersion: networking.k8s.io/v1
kind: NetworkPolicy
metadata:
  name: multi-port-egress
  namespace: default
spec:
  podSelector:
    matchLabels:
      role: db
  policyTypes:
  - Egress
  egress:
  - to:
    - ipBlock:
        cidr: 10.0.0.0/24
    ports:
    - protocol: TCP
      port: 32000
      endPort: 32768

The above rule allows any Pod with label role=db on the namespace default to communicate with any IP within the range 10.0.0.0/24 over TCP, provided that the target port is between the range 32000 and 32768.

The following restrictions apply when using this field:

  • The endPort field must be equal to or greater than the port field.
  • endPort can only be defined if port is also defined.
  • Both ports must be numeric.

Targeting a Namespace by its name

FEATURE STATE: Kubernetes 1.22 [stable]

The Kubernetes control plane sets an immutable label kubernetes.io/metadata.name on all namespaces, provided that the NamespaceDefaultLabelName feature gate is enabled. The value of the label is the namespace name.

While NetworkPolicy cannot target a namespace by its name with some object field, you can use the standardized label to target a specific namespace.

What you can't do with network policies (at least, not yet)

As of Kubernetes 1.25, the following functionality does not exist in the NetworkPolicy API, but you might be able to implement workarounds using Operating System components (such as SELinux, OpenVSwitch, IPTables, and so on) or Layer 7 technologies (Ingress controllers, Service Mesh implementations) or admission controllers. In case you are new to network security in Kubernetes, its worth noting that the following User Stories cannot (yet) be implemented using the NetworkPolicy API.

  • Forcing internal cluster traffic to go through a common gateway (this might be best served with a service mesh or other proxy).
  • Anything TLS related (use a service mesh or ingress controller for this).
  • Node specific policies (you can use CIDR notation for these, but you cannot target nodes by their Kubernetes identities specifically).
  • Targeting of services by name (you can, however, target pods or namespaces by their labels, which is often a viable workaround).
  • Creation or management of "Policy requests" that are fulfilled by a third party.
  • Default policies which are applied to all namespaces or pods (there are some third party Kubernetes distributions and projects which can do this).
  • Advanced policy querying and reachability tooling.
  • The ability to log network security events (for example connections that are blocked or accepted).
  • The ability to explicitly deny policies (currently the model for NetworkPolicies are deny by default, with only the ability to add allow rules).
  • The ability to prevent loopback or incoming host traffic (Pods cannot currently block localhost access, nor do they have the ability to block access from their resident node).

What's next

6 - DNS for Services and Pods

Your workload can discover Services within your cluster using DNS; this page explains how that works.

Kubernetes creates DNS records for Services and Pods. You can contact Services with consistent DNS names instead of IP addresses.

Kubernetes DNS schedules a DNS Pod and Service on the cluster, and configures the kubelets to tell individual containers to use the DNS Service's IP to resolve DNS names.

Every Service defined in the cluster (including the DNS server itself) is assigned a DNS name. By default, a client Pod's DNS search list includes the Pod's own namespace and the cluster's default domain.

Namespaces of Services

A DNS query may return different results based on the namespace of the Pod making it. DNS queries that don't specify a namespace are limited to the Pod's namespace. Access Services in other namespaces by specifying it in the DNS query.

For example, consider a Pod in a test namespace. A data Service is in the prod namespace.

A query for data returns no results, because it uses the Pod's test namespace.

A query for data.prod returns the intended result, because it specifies the namespace.

DNS queries may be expanded using the Pod's /etc/resolv.conf. Kubelet sets this file for each Pod. For example, a query for just data may be expanded to data.test.svc.cluster.local. The values of the search option are used to expand queries. To learn more about DNS queries, see the resolv.conf manual page.

nameserver 10.32.0.10
search <namespace>.svc.cluster.local svc.cluster.local cluster.local
options ndots:5

In summary, a Pod in the test namespace can successfully resolve either data.prod or data.prod.svc.cluster.local.

DNS Records

What objects get DNS records?

  1. Services
  2. Pods

The following sections detail the supported DNS record types and layout that is supported. Any other layout or names or queries that happen to work are considered implementation details and are subject to change without warning. For more up-to-date specification, see Kubernetes DNS-Based Service Discovery.

Services

A/AAAA records

"Normal" (not headless) Services are assigned a DNS A or AAAA record, depending on the IP family of the Service, for a name of the form my-svc.my-namespace.svc.cluster-domain.example. This resolves to the cluster IP of the Service.

"Headless" (without a cluster IP) Services are also assigned a DNS A or AAAA record, depending on the IP family of the Service, for a name of the form my-svc.my-namespace.svc.cluster-domain.example. Unlike normal Services, this resolves to the set of IPs of the Pods selected by the Service. Clients are expected to consume the set or else use standard round-robin selection from the set.

SRV records

SRV Records are created for named ports that are part of normal or Headless Services. For each named port, the SRV record would have the form _my-port-name._my-port-protocol.my-svc.my-namespace.svc.cluster-domain.example. For a regular Service, this resolves to the port number and the domain name: my-svc.my-namespace.svc.cluster-domain.example. For a headless Service, this resolves to multiple answers, one for each Pod that is backing the Service, and contains the port number and the domain name of the Pod of the form auto-generated-name.my-svc.my-namespace.svc.cluster-domain.example.

Pods

A/AAAA records

In general a Pod has the following DNS resolution:

pod-ip-address.my-namespace.pod.cluster-domain.example.

For example, if a Pod in the default namespace has the IP address 172.17.0.3, and the domain name for your cluster is cluster.local, then the Pod has a DNS name:

172-17-0-3.default.pod.cluster.local.

Any Pods exposed by a Service have the following DNS resolution available:

pod-ip-address.service-name.my-namespace.svc.cluster-domain.example.

Pod's hostname and subdomain fields

Currently when a Pod is created, its hostname is the Pod's metadata.name value.

The Pod spec has an optional hostname field, which can be used to specify the Pod's hostname. When specified, it takes precedence over the Pod's name to be the hostname of the Pod. For example, given a Pod with hostname set to "my-host", the Pod will have its hostname set to "my-host".

The Pod spec also has an optional subdomain field which can be used to specify its subdomain. For example, a Pod with hostname set to "foo", and subdomain set to "bar", in namespace "my-namespace", will have the fully qualified domain name (FQDN) "foo.bar.my-namespace.svc.cluster-domain.example".

Example:

apiVersion: v1
kind: Service
metadata:
  name: default-subdomain
spec:
  selector:
    name: busybox
  clusterIP: None
  ports:
  - name: foo # Actually, no port is needed.
    port: 1234
    targetPort: 1234
---
apiVersion: v1
kind: Pod
metadata:
  name: busybox1
  labels:
    name: busybox
spec:
  hostname: busybox-1
  subdomain: default-subdomain
  containers:
  - image: busybox:1.28
    command:
      - sleep
      - "3600"
    name: busybox
---
apiVersion: v1
kind: Pod
metadata:
  name: busybox2
  labels:
    name: busybox
spec:
  hostname: busybox-2
  subdomain: default-subdomain
  containers:
  - image: busybox:1.28
    command:
      - sleep
      - "3600"
    name: busybox

If there exists a headless Service in the same namespace as the Pod and with the same name as the subdomain, the cluster's DNS Server also returns an A or AAAA record for the Pod's fully qualified hostname. For example, given a Pod with the hostname set to "busybox-1" and the subdomain set to "default-subdomain", and a headless Service named "default-subdomain" in the same namespace, the Pod will see its own FQDN as "busybox-1.default-subdomain.my-namespace.svc.cluster-domain.example". DNS serves an A or AAAA record at that name, pointing to the Pod's IP. Both Pods "busybox1" and "busybox2" can have their distinct A or AAAA records.

An EndpointSlice can specify the DNS hostname for any endpoint addresses, along with its IP.

Pod's setHostnameAsFQDN field

FEATURE STATE: Kubernetes v1.22 [stable]

When a Pod is configured to have fully qualified domain name (FQDN), its hostname is the short hostname. For example, if you have a Pod with the fully qualified domain name busybox-1.default-subdomain.my-namespace.svc.cluster-domain.example, then by default the hostname command inside that Pod returns busybox-1 and the hostname --fqdn command returns the FQDN.

When you set setHostnameAsFQDN: true in the Pod spec, the kubelet writes the Pod's FQDN into the hostname for that Pod's namespace. In this case, both hostname and hostname --fqdn return the Pod's FQDN.

Pod's DNS Policy

DNS policies can be set on a per-Pod basis. Currently Kubernetes supports the following Pod-specific DNS policies. These policies are specified in the dnsPolicy field of a Pod Spec.

  • "Default": The Pod inherits the name resolution configuration from the node that the Pods run on. See related discussion for more details.
  • "ClusterFirst": Any DNS query that does not match the configured cluster domain suffix, such as "www.kubernetes.io", is forwarded to an upstream nameserver by the DNS server. Cluster administrators may have extra stub-domain and upstream DNS servers configured. See related discussion for details on how DNS queries are handled in those cases.
  • "ClusterFirstWithHostNet": For Pods running with hostNetwork, you should explicitly set its DNS policy to "ClusterFirstWithHostNet". Otherwise, Pods running with hostNetwork and "ClusterFirst" will fallback to the behavior of the "Default" policy.
    • Note: This is not supported on Windows. See below for details
  • "None": It allows a Pod to ignore DNS settings from the Kubernetes environment. All DNS settings are supposed to be provided using the dnsConfig field in the Pod Spec. See Pod's DNS config subsection below.

The example below shows a Pod with its DNS policy set to "ClusterFirstWithHostNet" because it has hostNetwork set to true.

apiVersion: v1
kind: Pod
metadata:
  name: busybox
  namespace: default
spec:
  containers:
  - image: busybox:1.28
    command:
      - sleep
      - "3600"
    imagePullPolicy: IfNotPresent
    name: busybox
  restartPolicy: Always
  hostNetwork: true
  dnsPolicy: ClusterFirstWithHostNet

Pod's DNS Config

FEATURE STATE: Kubernetes v1.14 [stable]

Pod's DNS Config allows users more control on the DNS settings for a Pod.

The dnsConfig field is optional and it can work with any dnsPolicy settings. However, when a Pod's dnsPolicy is set to "None", the dnsConfig field has to be specified.

Below are the properties a user can specify in the dnsConfig field:

  • nameservers: a list of IP addresses that will be used as DNS servers for the Pod. There can be at most 3 IP addresses specified. When the Pod's dnsPolicy is set to "None", the list must contain at least one IP address, otherwise this property is optional. The servers listed will be combined to the base nameservers generated from the specified DNS policy with duplicate addresses removed.
  • searches: a list of DNS search domains for hostname lookup in the Pod. This property is optional. When specified, the provided list will be merged into the base search domain names generated from the chosen DNS policy. Duplicate domain names are removed. Kubernetes allows for at most 6 search domains.
  • options: an optional list of objects where each object may have a name property (required) and a value property (optional). The contents in this property will be merged to the options generated from the specified DNS policy. Duplicate entries are removed.

The following is an example Pod with custom DNS settings:

apiVersion: v1
kind: Pod
metadata:
  namespace: default
  name: dns-example
spec:
  containers:
    - name: test
      image: nginx
  dnsPolicy: "None"
  dnsConfig:
    nameservers:
      - 1.2.3.4
    searches:
      - ns1.svc.cluster-domain.example
      - my.dns.search.suffix
    options:
      - name: ndots
        value: "2"
      - name: edns0

When the Pod above is created, the container test gets the following contents in its /etc/resolv.conf file:

nameserver 1.2.3.4
search ns1.svc.cluster-domain.example my.dns.search.suffix
options ndots:2 edns0

For IPv6 setup, search path and name server should be set up like this:

kubectl exec -it dns-example -- cat /etc/resolv.conf

The output is similar to this:

nameserver 2001:db8:30::a
search default.svc.cluster-domain.example svc.cluster-domain.example cluster-domain.example
options ndots:5

Expanded DNS Configuration

FEATURE STATE: Kubernetes 1.22 [alpha]

By default, for Pod's DNS Config, Kubernetes allows at most 6 search domains and a list of search domains of up to 256 characters.

If the feature gate ExpandedDNSConfig is enabled for the kube-apiserver and the kubelet, it is allowed for Kubernetes to have at most 32 search domains and a list of search domains of up to 2048 characters.

DNS resolution on Windows nodes

  • ClusterFirstWithHostNet is not supported for Pods that run on Windows nodes. Windows treats all names with a . as a FQDN and skips FQDN resolution.
  • On Windows, there are multiple DNS resolvers that can be used. As these come with slightly different behaviors, using the Resolve-DNSName powershell cmdlet for name query resolutions is recommended.
  • On Linux, you have a DNS suffix list, which is used after resolution of a name as fully qualified has failed. On Windows, you can only have 1 DNS suffix, which is the DNS suffix associated with that Pod's namespace (example: mydns.svc.cluster.local). Windows can resolve FQDNs, Services, or network name which can be resolved with this single suffix. For example, a Pod spawned in the default namespace, will have the DNS suffix default.svc.cluster.local. Inside a Windows Pod, you can resolve both kubernetes.default.svc.cluster.local and kubernetes, but not the partially qualified names (kubernetes.default or kubernetes.default.svc).

What's next

For guidance on administering DNS configurations, check Configure DNS Service

7 - IPv4/IPv6 dual-stack

Kubernetes lets you configure single-stack IPv4 networking, single-stack IPv6 networking, or dual stack networking with both network families active. This page explains how.
FEATURE STATE: Kubernetes v1.23 [stable]

IPv4/IPv6 dual-stack networking enables the allocation of both IPv4 and IPv6 addresses to Pods and Services.

IPv4/IPv6 dual-stack networking is enabled by default for your Kubernetes cluster starting in 1.21, allowing the simultaneous assignment of both IPv4 and IPv6 addresses.

Supported Features

IPv4/IPv6 dual-stack on your Kubernetes cluster provides the following features:

  • Dual-stack Pod networking (a single IPv4 and IPv6 address assignment per Pod)
  • IPv4 and IPv6 enabled Services
  • Pod off-cluster egress routing (eg. the Internet) via both IPv4 and IPv6 interfaces

Prerequisites

The following prerequisites are needed in order to utilize IPv4/IPv6 dual-stack Kubernetes clusters:

  • Kubernetes 1.20 or later

    For information about using dual-stack services with earlier Kubernetes versions, refer to the documentation for that version of Kubernetes.

  • Provider support for dual-stack networking (Cloud provider or otherwise must be able to provide Kubernetes nodes with routable IPv4/IPv6 network interfaces)

  • A network plugin that supports dual-stack networking.

Configure IPv4/IPv6 dual-stack

To configure IPv4/IPv6 dual-stack, set dual-stack cluster network assignments:

  • kube-apiserver:
    • --service-cluster-ip-range=<IPv4 CIDR>,<IPv6 CIDR>
  • kube-controller-manager:
    • --cluster-cidr=<IPv4 CIDR>,<IPv6 CIDR>
    • --service-cluster-ip-range=<IPv4 CIDR>,<IPv6 CIDR>
    • --node-cidr-mask-size-ipv4|--node-cidr-mask-size-ipv6 defaults to /24 for IPv4 and /64 for IPv6
  • kube-proxy:
    • --cluster-cidr=<IPv4 CIDR>,<IPv6 CIDR>
  • kubelet:
    • when there is no --cloud-provider the administrator can pass a comma-separated pair of IP addresses via --node-ip to manually configure dual-stack .status.addresses for that Node. If a Pod runs on that node in HostNetwork mode, the Pod reports these IP addresses in its .status.podIPs field. All podIPs in a node match the IP family preference defined by the .status.addresses field for that Node.

Services

You can create Services which can use IPv4, IPv6, or both.

The address family of a Service defaults to the address family of the first service cluster IP range (configured via the --service-cluster-ip-range flag to the kube-apiserver).

When you define a Service you can optionally configure it as dual stack. To specify the behavior you want, you set the .spec.ipFamilyPolicy field to one of the following values:

  • SingleStack: Single-stack service. The control plane allocates a cluster IP for the Service, using the first configured service cluster IP range.
  • PreferDualStack:
    • Allocates IPv4 and IPv6 cluster IPs for the Service.
  • RequireDualStack: Allocates Service .spec.ClusterIPs from both IPv4 and IPv6 address ranges.
    • Selects the .spec.ClusterIP from the list of .spec.ClusterIPs based on the address family of the first element in the .spec.ipFamilies array.

If you would like to define which IP family to use for single stack or define the order of IP families for dual-stack, you can choose the address families by setting an optional field, .spec.ipFamilies, on the Service.

You can set .spec.ipFamilies to any of the following array values:

  • ["IPv4"]
  • ["IPv6"]
  • ["IPv4","IPv6"] (dual stack)
  • ["IPv6","IPv4"] (dual stack)

The first family you list is used for the legacy .spec.ClusterIP field.

Dual-stack Service configuration scenarios

These examples demonstrate the behavior of various dual-stack Service configuration scenarios.

Dual-stack options on new Services

  1. This Service specification does not explicitly define .spec.ipFamilyPolicy. When you create this Service, Kubernetes assigns a cluster IP for the Service from the first configured service-cluster-ip-range and sets the .spec.ipFamilyPolicy to SingleStack. (Services without selectors and headless Services with selectors will behave in this same way.)

    apiVersion: v1
       kind: Service
       metadata:
         name: my-service
         labels:
           app.kubernetes.io/name: MyApp
       spec:
         selector:
           app.kubernetes.io/name: MyApp
         ports:
           - protocol: TCP
             port: 80
       
  2. This Service specification explicitly defines PreferDualStack in .spec.ipFamilyPolicy. When you create this Service on a dual-stack cluster, Kubernetes assigns both IPv4 and IPv6 addresses for the service. The control plane updates the .spec for the Service to record the IP address assignments. The field .spec.ClusterIPs is the primary field, and contains both assigned IP addresses; .spec.ClusterIP is a secondary field with its value calculated from .spec.ClusterIPs.

    • For the .spec.ClusterIP field, the control plane records the IP address that is from the same address family as the first service cluster IP range.
    • On a single-stack cluster, the .spec.ClusterIPs and .spec.ClusterIP fields both only list one address.
    • On a cluster with dual-stack enabled, specifying RequireDualStack in .spec.ipFamilyPolicy behaves the same as PreferDualStack.
    apiVersion: v1
       kind: Service
       metadata:
         name: my-service
         labels:
           app.kubernetes.io/name: MyApp
       spec:
         ipFamilyPolicy: PreferDualStack
         selector:
           app.kubernetes.io/name: MyApp
         ports:
           - protocol: TCP
             port: 80
       
  3. This Service specification explicitly defines IPv6 and IPv4 in .spec.ipFamilies as well as defining PreferDualStack in .spec.ipFamilyPolicy. When Kubernetes assigns an IPv6 and IPv4 address in .spec.ClusterIPs, .spec.ClusterIP is set to the IPv6 address because that is the first element in the .spec.ClusterIPs array, overriding the default.

    apiVersion: v1
       kind: Service
       metadata:
         name: my-service
         labels:
           app.kubernetes.io/name: MyApp
       spec:
         ipFamilyPolicy: PreferDualStack
         ipFamilies:
         - IPv6
         - IPv4
         selector:
           app.kubernetes.io/name: MyApp
         ports:
           - protocol: TCP
             port: 80
       

Dual-stack defaults on existing Services

These examples demonstrate the default behavior when dual-stack is newly enabled on a cluster where Services already exist. (Upgrading an existing cluster to 1.21 or beyond will enable dual-stack.)

  1. When dual-stack is enabled on a cluster, existing Services (whether IPv4 or IPv6) are configured by the control plane to set .spec.ipFamilyPolicy to SingleStack and set .spec.ipFamilies to the address family of the existing Service. The existing Service cluster IP will be stored in .spec.ClusterIPs.

    apiVersion: v1
       kind: Service
       metadata:
         name: my-service
         labels:
           app.kubernetes.io/name: MyApp
       spec:
         selector:
           app.kubernetes.io/name: MyApp
         ports:
           - protocol: TCP
             port: 80
       

    You can validate this behavior by using kubectl to inspect an existing service.

    kubectl get svc my-service -o yaml
    
    apiVersion: v1
    kind: Service
    metadata:
      labels:
        app.kubernetes.io/name: MyApp
      name: my-service
    spec:
      clusterIP: 10.0.197.123
      clusterIPs:
      - 10.0.197.123
      ipFamilies:
      - IPv4
      ipFamilyPolicy: SingleStack
      ports:
      - port: 80
        protocol: TCP
        targetPort: 80
      selector:
        app.kubernetes.io/name: MyApp
      type: ClusterIP
    status:
      loadBalancer: {}
    
  2. When dual-stack is enabled on a cluster, existing headless Services with selectors are configured by the control plane to set .spec.ipFamilyPolicy to SingleStack and set .spec.ipFamilies to the address family of the first service cluster IP range (configured via the --service-cluster-ip-range flag to the kube-apiserver) even though .spec.ClusterIP is set to None.

    apiVersion: v1
       kind: Service
       metadata:
         name: my-service
         labels:
           app.kubernetes.io/name: MyApp
       spec:
         selector:
           app.kubernetes.io/name: MyApp
         ports:
           - protocol: TCP
             port: 80
       

    You can validate this behavior by using kubectl to inspect an existing headless service with selectors.

    kubectl get svc my-service -o yaml
    
    apiVersion: v1
    kind: Service
    metadata:
      labels:
        app.kubernetes.io/name: MyApp
      name: my-service
    spec:
      clusterIP: None
      clusterIPs:
      - None
      ipFamilies:
      - IPv4
      ipFamilyPolicy: SingleStack
      ports:
      - port: 80
        protocol: TCP
        targetPort: 80
      selector:
        app.kubernetes.io/name: MyApp
    

Switching Services between single-stack and dual-stack

Services can be changed from single-stack to dual-stack and from dual-stack to single-stack.

  1. To change a Service from single-stack to dual-stack, change .spec.ipFamilyPolicy from SingleStack to PreferDualStack or RequireDualStack as desired. When you change this Service from single-stack to dual-stack, Kubernetes assigns the missing address family so that the Service now has IPv4 and IPv6 addresses.

    Edit the Service specification updating the .spec.ipFamilyPolicy from SingleStack to PreferDualStack.

    Before:

    spec:
      ipFamilyPolicy: SingleStack
    

    After:

    spec:
      ipFamilyPolicy: PreferDualStack
    
  2. To change a Service from dual-stack to single-stack, change .spec.ipFamilyPolicy from PreferDualStack or RequireDualStack to SingleStack. When you change this Service from dual-stack to single-stack, Kubernetes retains only the first element in the .spec.ClusterIPs array, and sets .spec.ClusterIP to that IP address and sets .spec.ipFamilies to the address family of .spec.ClusterIPs.

Headless Services without selector

For Headless Services without selectors and without .spec.ipFamilyPolicy explicitly set, the .spec.ipFamilyPolicy field defaults to RequireDualStack.

Service type LoadBalancer

To provision a dual-stack load balancer for your Service:

  • Set the .spec.type field to LoadBalancer
  • Set .spec.ipFamilyPolicy field to PreferDualStack or RequireDualStack

Egress traffic

If you want to enable egress traffic in order to reach off-cluster destinations (eg. the public Internet) from a Pod that uses non-publicly routable IPv6 addresses, you need to enable the Pod to use a publicly routed IPv6 address via a mechanism such as transparent proxying or IP masquerading. The ip-masq-agent project supports IP masquerading on dual-stack clusters.

Windows support

Kubernetes on Windows does not support single-stack "IPv6-only" networking. However, dual-stack IPv4/IPv6 networking for pods and nodes with single-family services is supported.

You can use IPv4/IPv6 dual-stack networking with l2bridge networks.

You can read more about the different network modes for Windows within the Networking on Windows topic.

What's next

8 - Topology Aware Hints

Topology Aware Hints provides a mechanism to help keep network traffic within the zone where it originated. Preferring same-zone traffic between Pods in your cluster can help with reliability, performance (network latency and throughput), or cost.
FEATURE STATE: Kubernetes v1.23 [beta]

Topology Aware Hints enable topology aware routing by including suggestions for how clients should consume endpoints. This approach adds metadata to enable consumers of EndpointSlice (or Endpoints) objects, so that traffic to those network endpoints can be routed closer to where it originated.

For example, you can route traffic within a locality to reduce costs, or to improve network performance.

Motivation

Kubernetes clusters are increasingly deployed in multi-zone environments. Topology Aware Hints provides a mechanism to help keep traffic within the zone it originated from. This concept is commonly referred to as "Topology Aware Routing". When calculating the endpoints for a Service, the EndpointSlice controller considers the topology (region and zone) of each endpoint and populates the hints field to allocate it to a zone. Cluster components such as the kube-proxy can then consume those hints, and use them to influence how the traffic is routed (favoring topologically closer endpoints).

Using Topology Aware Hints

You can activate Topology Aware Hints for a Service by setting the service.kubernetes.io/topology-aware-hints annotation to auto. This tells the EndpointSlice controller to set topology hints if it is deemed safe. Importantly, this does not guarantee that hints will always be set.

How it works

The functionality enabling this feature is split into two components: The EndpointSlice controller and the kube-proxy. This section provides a high level overview of how each component implements this feature.

EndpointSlice controller

The EndpointSlice controller is responsible for setting hints on EndpointSlices when this feature is enabled. The controller allocates a proportional amount of endpoints to each zone. This proportion is based on the allocatable CPU cores for nodes running in that zone. For example, if one zone had 2 CPU cores and another zone only had 1 CPU core, the controller would allocate twice as many endpoints to the zone with 2 CPU cores.

The following example shows what an EndpointSlice looks like when hints have been populated:

apiVersion: discovery.k8s.io/v1
kind: EndpointSlice
metadata:
  name: example-hints
  labels:
    kubernetes.io/service-name: example-svc
addressType: IPv4
ports:
  - name: http
    protocol: TCP
    port: 80
endpoints:
  - addresses:
      - "10.1.2.3"
    conditions:
      ready: true
    hostname: pod-1
    zone: zone-a
    hints:
      forZones:
        - name: "zone-a"

kube-proxy

The kube-proxy component filters the endpoints it routes to based on the hints set by the EndpointSlice controller. In most cases, this means that the kube-proxy is able to route traffic to endpoints in the same zone. Sometimes the controller allocates endpoints from a different zone to ensure more even distribution of endpoints between zones. This would result in some traffic being routed to other zones.

Safeguards

The Kubernetes control plane and the kube-proxy on each node apply some safeguard rules before using Topology Aware Hints. If these don't check out, the kube-proxy selects endpoints from anywhere in your cluster, regardless of the zone.

  1. Insufficient number of endpoints: If there are less endpoints than zones in a cluster, the controller will not assign any hints.

  2. Impossible to achieve balanced allocation: In some cases, it will be impossible to achieve a balanced allocation of endpoints among zones. For example, if zone-a is twice as large as zone-b, but there are only 2 endpoints, an endpoint allocated to zone-a may receive twice as much traffic as zone-b. The controller does not assign hints if it can't get this "expected overload" value below an acceptable threshold for each zone. Importantly this is not based on real-time feedback. It is still possible for individual endpoints to become overloaded.

  3. One or more Nodes has insufficient information: If any node does not have a topology.kubernetes.io/zone label or is not reporting a value for allocatable CPU, the control plane does not set any topology-aware endpoint hints and so kube-proxy does not filter endpoints by zone.

  4. One or more endpoints does not have a zone hint: When this happens, the kube-proxy assumes that a transition from or to Topology Aware Hints is underway. Filtering endpoints for a Service in this state would be dangerous so the kube-proxy falls back to using all endpoints.

  5. A zone is not represented in hints: If the kube-proxy is unable to find at least one endpoint with a hint targeting the zone it is running in, it falls to using endpoints from all zones. This is most likely to happen as you add a new zone into your existing cluster.

Constraints

  • Topology Aware Hints are not used when either externalTrafficPolicy or internalTrafficPolicy is set to Local on a Service. It is possible to use both features in the same cluster on different Services, just not on the same Service.

  • This approach will not work well for Services that have a large proportion of traffic originating from a subset of zones. Instead this assumes that incoming traffic will be roughly proportional to the capacity of the Nodes in each zone.

  • The EndpointSlice controller ignores unready nodes as it calculates the proportions of each zone. This could have unintended consequences if a large portion of nodes are unready.

  • The EndpointSlice controller does not take into account tolerations when deploying or calculating the proportions of each zone. If the Pods backing a Service are limited to a subset of Nodes in the cluster, this will not be taken into account.

  • This may not work well with autoscaling. For example, if a lot of traffic is originating from a single zone, only the endpoints allocated to that zone will be handling that traffic. That could result in Horizontal Pod Autoscaler either not picking up on this event, or newly added pods starting in a different zone.

What's next

9 - Networking on Windows

Kubernetes supports running nodes on either Linux or Windows. You can mix both kinds of node within a single cluster. This page provides an overview to networking specific to the Windows operating system.

Container networking on Windows

Networking for Windows containers is exposed through CNI plugins. Windows containers function similarly to virtual machines in regards to networking. Each container has a virtual network adapter (vNIC) which is connected to a Hyper-V virtual switch (vSwitch). The Host Networking Service (HNS) and the Host Compute Service (HCS) work together to create containers and attach container vNICs to networks. HCS is responsible for the management of containers whereas HNS is responsible for the management of networking resources such as:

  • Virtual networks (including creation of vSwitches)
  • Endpoints / vNICs
  • Namespaces
  • Policies including packet encapsulations, load-balancing rules, ACLs, and NAT rules.

The Windows HNS and vSwitch implement namespacing and can create virtual NICs as needed for a pod or container. However, many configurations such as DNS, routes, and metrics are stored in the Windows registry database rather than as files inside /etc, which is how Linux stores those configurations. The Windows registry for the container is separate from that of the host, so concepts like mapping /etc/resolv.conf from the host into a container don't have the same effect they would on Linux. These must be configured using Windows APIs run in the context of that container. Therefore CNI implementations need to call the HNS instead of relying on file mappings to pass network details into the pod or container.

Network modes

Windows supports five different networking drivers/modes: L2bridge, L2tunnel, Overlay (Beta), Transparent, and NAT. In a heterogeneous cluster with Windows and Linux worker nodes, you need to select a networking solution that is compatible on both Windows and Linux. The following table lists the out-of-tree plugins are supported on Windows, with recommendations on when to use each CNI:

Network Driver Description Container Packet Modifications Network Plugins Network Plugin Characteristics
L2bridge Containers are attached to an external vSwitch. Containers are attached to the underlay network, although the physical network doesn't need to learn the container MACs because they are rewritten on ingress/egress. MAC is rewritten to host MAC, IP may be rewritten to host IP using HNS OutboundNAT policy. win-bridge, Azure-CNI, Flannel host-gateway uses win-bridge win-bridge uses L2bridge network mode, connects containers to the underlay of hosts, offering best performance. Requires user-defined routes (UDR) for inter-node connectivity.
L2Tunnel This is a special case of l2bridge, but only used on Azure. All packets are sent to the virtualization host where SDN policy is applied. MAC rewritten, IP visible on the underlay network Azure-CNI Azure-CNI allows integration of containers with Azure vNET, and allows them to leverage the set of capabilities that Azure Virtual Network provides. For example, securely connect to Azure services or use Azure NSGs. See azure-cni for some examples
Overlay Containers are given a vNIC connected to an external vSwitch. Each overlay network gets its own IP subnet, defined by a custom IP prefix.The overlay network driver uses VXLAN encapsulation. Encapsulated with an outer header. win-overlay, Flannel VXLAN (uses win-overlay) win-overlay should be used when virtual container networks are desired to be isolated from underlay of hosts (e.g. for security reasons). Allows for IPs to be re-used for different overlay networks (which have different VNID tags) if you are restricted on IPs in your datacenter. This option requires KB4489899 on Windows Server 2019.
Transparent (special use case for ovn-kubernetes) Requires an external vSwitch. Containers are attached to an external vSwitch which enables intra-pod communication via logical networks (logical switches and routers). Packet is encapsulated either via GENEVE or STT tunneling to reach pods which are not on the same host.
Packets are forwarded or dropped via the tunnel metadata information supplied by the ovn network controller.
NAT is done for north-south communication.
ovn-kubernetes Deploy via ansible. Distributed ACLs can be applied via Kubernetes policies. IPAM support. Load-balancing can be achieved without kube-proxy. NATing is done without using iptables/netsh.
NAT (not used in Kubernetes) Containers are given a vNIC connected to an internal vSwitch. DNS/DHCP is provided using an internal component called WinNAT MAC and IP is rewritten to host MAC/IP. nat Included here for completeness

As outlined above, the Flannel CNI plugin is also supported on Windows via the VXLAN network backend (Beta support ; delegates to win-overlay) and host-gateway network backend (stable support; delegates to win-bridge).

This plugin supports delegating to one of the reference CNI plugins (win-overlay, win-bridge), to work in conjunction with Flannel daemon on Windows (Flanneld) for automatic node subnet lease assignment and HNS network creation. This plugin reads in its own configuration file (cni.conf), and aggregates it with the environment variables from the FlannelD generated subnet.env file. It then delegates to one of the reference CNI plugins for network plumbing, and sends the correct configuration containing the node-assigned subnet to the IPAM plugin (for example: host-local).

For Node, Pod, and Service objects, the following network flows are supported for TCP/UDP traffic:

  • Pod → Pod (IP)
  • Pod → Pod (Name)
  • Pod → Service (Cluster IP)
  • Pod → Service (PQDN, but only if there are no ".")
  • Pod → Service (FQDN)
  • Pod → external (IP)
  • Pod → external (DNS)
  • Node → Pod
  • Pod → Node

IP address management (IPAM)

The following IPAM options are supported on Windows:

Load balancing and Services

A Kubernetes Service is an abstraction that defines a logical set of Pods and a means to access them over a network. In a cluster that includes Windows nodes, you can use the following types of Service:

  • NodePort
  • ClusterIP
  • LoadBalancer
  • ExternalName

Windows container networking differs in some important ways from Linux networking. The Microsoft documentation for Windows Container Networking provides additional details and background.

On Windows, you can use the following settings to configure Services and load balancing behavior:

Windows Service Settings
Feature Description Minimum Supported Windows OS build How to enable
Session affinity Ensures that connections from a particular client are passed to the same Pod each time. Windows Server 2022 Set service.spec.sessionAffinity to "ClientIP"
Direct Server Return (DSR) Load balancing mode where the IP address fixups and the LBNAT occurs at the container vSwitch port directly; service traffic arrives with the source IP set as the originating pod IP. Windows Server 2019 Set the following flags in kube-proxy: --feature-gates="WinDSR=true" --enable-dsr=true
Preserve-Destination Skips DNAT of service traffic, thereby preserving the virtual IP of the target service in packets reaching the backend Pod. Also disables node-node forwarding. Windows Server, version 1903 Set "preserve-destination": "true" in service annotations and enable DSR in kube-proxy.
IPv4/IPv6 dual-stack networking Native IPv4-to-IPv4 in parallel with IPv6-to-IPv6 communications to, from, and within a cluster Windows Server 2019 See IPv4/IPv6 dual-stack
Client IP preservation Ensures that source IP of incoming ingress traffic gets preserved. Also disables node-node forwarding. Windows Server 2019 Set service.spec.externalTrafficPolicy to "Local" and enable DSR in kube-proxy

Limitations

The following networking functionality is not supported on Windows nodes:

  • Host networking mode
  • Local NodePort access from the node itself (works for other nodes or external clients)
  • More than 64 backend pods (or unique destination addresses) for a single Service
  • IPv6 communication between Windows pods connected to overlay networks
  • Local Traffic Policy in non-DSR mode
  • Outbound communication using the ICMP protocol via the win-overlay, win-bridge, or using the Azure-CNI plugin.
    Specifically, the Windows data plane (VFP) doesn't support ICMP packet transpositions, and this means:
    • ICMP packets directed to destinations within the same network (such as pod to pod communication via ping) work as expected;
    • TCP/UDP packets work as expected;
    • ICMP packets directed to pass through a remote network (e.g. pod to external internet communication via ping) cannot be transposed and thus will not be routed back to their source;
    • Since TCP/UDP packets can still be transposed, you can substitute ping <destination> with curl <destination> when debugging connectivity with the outside world.

Other limitations:

  • Windows reference network plugins win-bridge and win-overlay do not implement CNI spec v0.4.0, due to a missing CHECK implementation.
  • The Flannel VXLAN CNI plugin has the following limitations on Windows:
    • Node-pod connectivity is only possible for local pods with Flannel v0.12.0 (or higher).
    • Flannel is restricted to using VNI 4096 and UDP port 4789. See the official Flannel VXLAN backend docs for more details on these parameters.

10 - Service Internal Traffic Policy

If two Pods in your cluster want to communicate, and both Pods are actually running on the same node, use Service Internal Traffic Policy to keep network traffic within that node. Avoiding a round trip via the cluster network can help with reliability, performance (network latency and throughput), or cost.
FEATURE STATE: Kubernetes v1.23 [beta]

Service Internal Traffic Policy enables internal traffic restrictions to only route internal traffic to endpoints within the node the traffic originated from. The "internal" traffic here refers to traffic originated from Pods in the current cluster. This can help to reduce costs and improve performance.

Using Service Internal Traffic Policy

The ServiceInternalTrafficPolicy feature gate is a Beta feature and enabled by default. When the feature is enabled, you can enable the internal-only traffic policy for a Service, by setting its .spec.internalTrafficPolicy to Local. This tells kube-proxy to only use node local endpoints for cluster internal traffic.

The following example shows what a Service looks like when you set .spec.internalTrafficPolicy to Local:

apiVersion: v1
kind: Service
metadata:
  name: my-service
spec:
  selector:
    app.kubernetes.io/name: MyApp
  ports:
    - protocol: TCP
      port: 80
      targetPort: 9376
  internalTrafficPolicy: Local

How it works

The kube-proxy filters the endpoints it routes to based on the spec.internalTrafficPolicy setting. When it's set to Local, only node local endpoints are considered. When it's Cluster or missing, all endpoints are considered. When the feature gate ServiceInternalTrafficPolicy is enabled, spec.internalTrafficPolicy defaults to "Cluster".

What's next

11 - Topology-aware traffic routing with topology keys

FEATURE STATE: Kubernetes v1.21 [deprecated]

Service Topology enables a service to route traffic based upon the Node topology of the cluster. For example, a service can specify that traffic be preferentially routed to endpoints that are on the same Node as the client, or in the same availability zone.

Topology-aware traffic routing

By default, traffic sent to a ClusterIP or NodePort Service may be routed to any backend address for the Service. Kubernetes 1.7 made it possible to route "external" traffic to the Pods running on the same Node that received the traffic. For ClusterIP Services, the equivalent same-node preference for routing wasn't possible; nor could you configure your cluster to favor routing to endpoints within the same zone. By setting topologyKeys on a Service, you're able to define a policy for routing traffic based upon the Node labels for the originating and destination Nodes.

The label matching between the source and destination lets you, as a cluster operator, designate sets of Nodes that are "closer" and "farther" from one another. You can define labels to represent whatever metric makes sense for your own requirements. In public clouds, for example, you might prefer to keep network traffic within the same zone, because interzonal traffic has a cost associated with it (and intrazonal traffic typically does not). Other common needs include being able to route traffic to a local Pod managed by a DaemonSet, or directing traffic to Nodes connected to the same top-of-rack switch for the lowest latency.

Using Service Topology

If your cluster has the ServiceTopology feature gate enabled, you can control Service traffic routing by specifying the topologyKeys field on the Service spec. This field is a preference-order list of Node labels which will be used to sort endpoints when accessing this Service. Traffic will be directed to a Node whose value for the first label matches the originating Node's value for that label. If there is no backend for the Service on a matching Node, then the second label will be considered, and so forth, until no labels remain.

If no match is found, the traffic will be rejected, as if there were no backends for the Service at all. That is, endpoints are chosen based on the first topology key with available backends. If this field is specified and all entries have no backends that match the topology of the client, the service has no backends for that client and connections should fail. The special value "*" may be used to mean "any topology". This catch-all value, if used, only makes sense as the last value in the list.

If topologyKeys is not specified or empty, no topology constraints will be applied.

Consider a cluster with Nodes that are labeled with their hostname, zone name, and region name. Then you can set the topologyKeys values of a service to direct traffic as follows.

  • Only to endpoints on the same node, failing if no endpoint exists on the node: ["kubernetes.io/hostname"].
  • Preferentially to endpoints on the same node, falling back to endpoints in the same zone, followed by the same region, and failing otherwise: ["kubernetes.io/hostname", "topology.kubernetes.io/zone", "topology.kubernetes.io/region"]. This may be useful, for example, in cases where data locality is critical.
  • Preferentially to the same zone, but fallback on any available endpoint if none are available within this zone: ["topology.kubernetes.io/zone", "*"].

Constraints

  • Service topology is not compatible with externalTrafficPolicy=Local, and therefore a Service cannot use both of these features. It is possible to use both features in the same cluster on different Services, only not on the same Service.

  • Valid topology keys are currently limited to kubernetes.io/hostname, topology.kubernetes.io/zone, and topology.kubernetes.io/region, but will be generalized to other node labels in the future.

  • Topology keys must be valid label keys and at most 16 keys may be specified.

  • The catch-all value, "*", must be the last value in the topology keys, if it is used.

Examples

The following are common examples of using the Service Topology feature.

Only Node Local Endpoints

A Service that only routes to node local endpoints. If no endpoints exist on the node, traffic is dropped:

apiVersion: v1
kind: Service
metadata:
  name: my-service
spec:
  selector:
    app: my-app
  ports:
    - protocol: TCP
      port: 80
      targetPort: 9376
  topologyKeys:
    - "kubernetes.io/hostname"

Prefer Node Local Endpoints

A Service that prefers node local Endpoints but falls back to cluster wide endpoints if node local endpoints do not exist:

apiVersion: v1
kind: Service
metadata:
  name: my-service
spec:
  selector:
    app: my-app
  ports:
    - protocol: TCP
      port: 80
      targetPort: 9376
  topologyKeys:
    - "kubernetes.io/hostname"
    - "*"

Only Zonal or Regional Endpoints

A Service that prefers zonal then regional endpoints. If no endpoints exist in either, traffic is dropped.

apiVersion: v1
kind: Service
metadata:
  name: my-service
spec:
  selector:
    app: my-app
  ports:
    - protocol: TCP
      port: 80
      targetPort: 9376
  topologyKeys:
    - "topology.kubernetes.io/zone"
    - "topology.kubernetes.io/region"

Prefer Node Local, Zonal, then Regional Endpoints

A Service that prefers node local, zonal, then regional endpoints but falls back to cluster wide endpoints.

apiVersion: v1
kind: Service
metadata:
  name: my-service
spec:
  selector:
    app: my-app
  ports:
    - protocol: TCP
      port: 80
      targetPort: 9376
  topologyKeys:
    - "kubernetes.io/hostname"
    - "topology.kubernetes.io/zone"
    - "topology.kubernetes.io/region"
    - "*"

What's next