EndpointSlices

EndpointSlices provide a simple way to track network endpoints within a Kubernetes cluster. They offer a more scalable and extensible alternative to Endpoints.

Motivation

The Endpoints API has provided a simple and straightforward way of tracking network endpoints in Kubernetes. Unfortunately as Kubernetes clusters and Services have grown to handle and send more traffic to more backend Pods, 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 resource, those resources could get quite large. That affected the performance of Kubernetes components (notably the master control plane) and resulted in significant amounts of network traffic and processing when Endpoints changed. EndpointSlices help you mitigate those issues as well as provide an extensible platform for additional features such as topological routing.

EndpointSlice resources

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 resource for the example Kubernetes Service.

apiVersion: discovery.k8s.io/v1beta1
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
    topology:
      kubernetes.io/hostname: node-1
      topology.kubernetes.io/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. When enabled, they should provide a performance improvement for services with large numbers of endpoints.

Address types

EndpointSlices support three address types:

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

Topology information

Each endpoint within an EndpointSlice can contain relevant topology information. This is used to indicate where an endpoint is, containing information about the corresponding Node, zone, and region. When the values are available, the control plane sets the following Topology labels for EndpointSlices:

• kubernetes.io/hostname - The name of the Node this endpoint is on.
• topology.kubernetes.io/zone - The zone this endpoint is in.
• topology.kubernetes.io/region - The region this endpoint is in.

The values of these labels are derived from resources associated with each endpoint in a slice. The hostname label represents the value of the NodeName field on the corresponding Pod. The zone and region labels represent the value of the labels with the same names on the corresponding Node.

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. Implementations using EndpointSlice must be able to have the endpoint appear in more than one slice. A reference implementation of how to perform endpoint deduplication can be found in the EndpointSliceCache implementation in kube-proxy.

What's next

• Learn about Enabling EndpointSlices
• Read Connecting Applications with Services