Antrea Multi-cluster Architecture

Antrea Multi-cluster implements Multi-cluster Service API, which allows users to create multi-cluster Services that can be accessed cross clusters in a ClusterSet. Antrea Multi-cluster also supports Antrea ClusterNetworkPolicy replication. Multi-cluster admins can define ClusterNetworkPolicies to be replicated across the entire ClusterSet, and enforced in all member clusters.

An Antrea Multi-cluster ClusterSet includes a leader cluster and multiple member clusters. Antrea Multi-cluster Controller needs to be deployed in the leader and all member clusters. A cluster can serve as the leader, and meanwhile also be a member cluster of the ClusterSet.

The diagram below depicts a basic Antrea Multi-cluster topology with one leader cluster and two member clusters.

Antrea Multi-cluster Topology


ClusterSet is a placeholder name for a group of clusters with a high degree of mutual trust and shared ownership that share Services amongst themselves. Within a ClusterSet, Namespace sameness applies, which means all Namespaces with a given name are considered to be the same Namespace. The ClusterSet Custom Resource Definition(CRD) defines a ClusterSet including the leader and member clusters information.

The ClusterClaim CRD is used to claim a ClusterSet with a unique ClusterSet ID, and to claim the cluster itself as a member of a ClusterSet with a unique cluster ID.

The MemberClusterAnnounce CRD declares a member cluster configuration to the leader cluster.

The Common Area is an abstraction in the Antrea Multi-cluster implementation provides a storage interface for resource export/import that can be read/written by all member and leader clusters in the ClusterSet. The Common Area is implemented with a Namespace in the leader cluster for a given ClusterSet.

Antrea Multi-cluster Controller

Antrea Multi-cluster Controller implements ClusterSet management and resource export/import in the ClusterSet. In either a leader or a member cluster, Antrea Multi-cluster Controller is deployed with a Deployment of a single replica, but it takes different responsibilities in leader and member clusters.

ClusterSet Establishment

In a member cluster, Multi-cluster Controller watches and validates the ClusterSet and ClusterClaim CRs, and creates a MemberClusterAnnounce CR in the Common Area of the leader cluster to join the ClusterSet.

In the leader cluster, Multi-cluster controller watches and validates the ClusterSet and Clusterclaim CRs, and initializes the ClusterSet. It also validates the MemberClusterAnnounce CR created by a member cluster and adds the cluster to the ClusterSet CR’s member cluster list.

Resource Export and Import

In a member cluster, Multi-cluster controller watches exported resources (e.g. ServiceExports, Services, Multi-cluster Gateways), encapsulates an exported resource into a ResourceExport and creates the ResourceExport CR in the Common Area of the leader cluster.

In the leader cluster, Multi-cluster Controller watches ResourceExports created by member clusters (in the case of Service and ClusterInfo export), or by the ClusterSet admin (in the case of Multi-cluster NetworkPolicy), converts ResourceExports to ResourceImports, and creates the ResourceImport CRs in the Common Area for member clusters to import them. Multi-cluster Controller also merges ResourceExports from different member clusters to a single ResourceImport, when these exported resources share the same kind, name, and original Namespace (matching Namespace sameness).

Multi-cluster Controller in a member cluster also watches ResourceImports in the Common Area of the leader cluster, decapsulates the resources from them, and creates the resources (e.g. Services, Endpoints, Antrea ClusterNetworkPolicies, ClusterInfoImports) in the member cluster.

For more information about multi-cluster Service export/import, please also check the Service Export and Import section.

Multi-cluster Service

Service Export and Import

Antrea Multi-cluster Service Export/Import Pipeline

Antrea Multi-cluster Controller implements Service export/import among member clusters. The above diagram depicts Antrea Multi-cluster resource export/import pipeline, using Service export/import as an example.

Given two Services with the same name and Namespace in two member clusters - foo.ns.cluster-a.local and foo.ns.cluster-b.local, a multi-cluster Service can be created by the following resource export/import workflow.

  • User creates a ServiceExport foo in Namespace ns in each of the two clusters.
  • Multi-cluster Controllers in cluster-a and cluster-b see ServiceExport foo, and both create two ResourceExports for the Service and Endpoints respectively in the Common Area of the leader cluster.
  • Multi-cluster Controller in the leader cluster sees the ResourcesExports in the Common Area, including the two for Service foo: cluster-a-ns-foo-service, cluster-b-ns-foo-service; and the two for the Endpoints: cluster-a-ns-foo-endpoints, cluster-b-ns-foo-endpoints. It then creates a ResourceImport ns-foo-service for the multi-cluster Service; and a ResourceImport ns-foo-endpoints for the Endpoints, which includes the exported endpoints of both cluster-a-ns-foo-endpoints and cluster-b-ns-foo-endpoints.
  • Multi-cluster Controller in each member cluster watches the ResourceImports from the Common Area, decapsulates them and gets Service ns/antrea-mc-foo and Endpoints ns/antrea-mc-foo, and creates the Service and Endpoints, as well as a ServiceImport foo in the local Namespace ns.

Service Access Across Clusters

Since Antrea v1.7.0, the Service’s ClusterIP is exported as the multi-cluster Service’s Endpoints. Multi-cluster Gateways must be configured to support multi-cluster Service access across member clusters, and Service CIDRs cannot overlap between clusters. Please refer to Multi-cluster Gateway for more information. Before Antrea v1.7.0, Pod IPs are exported as the multi-cluster Service’s Endpoints. Pod IPs must be directly reachable across clusters for multi-cluster Service access, and Pod CIDRs cannot overlap between clusters. Antrea Multi-cluster only supports creating multi-cluster Services for Services of type ClusterIP.

Multi-cluster Gateway

Antrea started to support Multi-cluster Gateway since v1.7.0. User can choose one K8s Node as the Multi-cluster Gateway in a member cluster. The Gateway Node is responsible for routing all cross-clusters traffic from the local cluster to other member clusters through tunnels. The diagram below depicts Antrea Multi-cluster connectivity with Multi-cluster Gateways.

Antrea Multi-cluster Gateway

Antrea Agent is responsible for setting up tunnels between Gateways of meamber clusters. At the moment, Multi-cluster Gateway only works with Antrea encap mode. The tunnels between Gateways use Antrea Agent’s configured tunnel type. All member clusters in a ClusterSet need to deploy Antrea with the same tunnel type.

The Multi-cluster Gateway implementation introduces two new CRDs Gateway and ClusterInfoImport. Gateway includes the local Multi-cluster Gateway information including: internalIP for tunnels to local Nodes, and gatewayIP for tunnels to remote cluster Gateways. ClusterInfoImport includes Gateway and network information of member clusters, including Gateway IPs and Service CIDRs. The existing esource export/import pipeline is leveraged to exchange the cluster network information among member clusters, generating ClusterInfoImports in each member cluster.

Multi-cluster Service Traffic Walk

Let’s use the ClusterSet in the above diagram as an example. As shown in the diagram:

  1. Cluster A has a client Pod named pod-a running on a regular Node, and a multi-cluster Service named antrea-mc-nginx with ClusterIP in the default Namespace.
  2. Cluster B exported a Service named nginx with ClusterIP in the default Namespace. The Service has one Endpoint which is pod-b’s IP.
  3. Cluster C exported a Service named nginx with ClusterIP also in the default Namespace. The Service has one Endpoint which is pod-c’s IP.

The multi-cluster Service antrea-mc-nginx in cluster A will have two Endpoints:

  • nginx Service’s ClusterIP from cluster B.
  • nginx Service’s ClusterIP from cluster C.

When the client Pod pod-a on cluster A tries to access the multi-cluster Service antrea-mc-nginx, the request packet will first go through the Service load balancing pipeline on the source Node node-a2, with one endpoint of the multi-cluster Service being chosen as the destination. Let’s say endpoint from cluster C is chosen, then the request packet will be DNAT’d with IP and tunnelled to the local Gateway Node node-a1. node-a1 knows from the destination IP ( the packet is multi-cluster Service traffic destined for cluster C, and it will tunnel the packet to cluster C’s Gateway Node node-c1, after performing SNAT and setting the packet’s source IP to its own Gateway IP. On node-c1, the packet will go through the Service load balancing pipeline again with an endpoint of Service nginx being chosen as the destination. As the Service has only one endpoint - of pod-c, the request packet will be DNAT’d to and tunnelled to node-c2 where pod-c is running. Finally, on node-c2 the packet will go through the normal Antrea forwarding pipeline and be forwarded to pod-c.

Antrea Multi-cluster NetworkPolicy

At this moment, Antrea does not support Pod-level policy enforcement for cross-cluster traffic. Access towards multi-cluster Services can be regulated with Antrea ClusterNetworkPolicy toService rules. In each member cluster, users can create an Antrea ClusterNetworkPolicy selecting Pods in that cluster, with the imported Mutli-cluster Service name and Namespace in an egress toService rule, and the Action to take for traffic matching this rule. For more information regarding Antrea ClusterNetworkPolicy (ACNP), refer to this document.

Multi-cluster admins can also specify certain ClusterNetworkPolicies to be replicated across the entire ClusterSet. The ACNP to be replicated should be created as a ResourceExport in the leader cluster, and the resource export/import pipeline will ensure member clusters receive this ACNP spec to be replicated. Each member cluster’s Multi-cluster Controller will then create an ACNP in their respective clusters.

Getting Started

To help you get started, see the documentation.