Set up a multi-cluster mesh on GKE

This guide explains how to join two clusters into a single Cloud Service Mesh using Mesh CA or Istio CA, and enable cross-cluster load balancing. You can easily extend this process to incorporate any number of clusters into your mesh.

A multi-cluster Cloud Service Mesh configuration can solve several crucial enterprise scenarios, such as scale, location, and isolation. For more information, see Multi-cluster use cases.

Prerequisites

This guide assumes that you have two or more Google Cloud GKE clusters that meet the following requirements:

• Cloud Service Mesh version 1.11 or later installed on the clusters using asmcli install. You need asmcli, the istioctl tool, and samples that asmcli downloads to the directory that you specified in --output_dir when you ran asmcli install If need to get set up, follow the steps in Install dependent tools and validate cluster to:

  • Install required tools
  • Download asmcli
  • Grant cluster admin permissions
  • Validate your project and cluster

• Clusters in your mesh must have connectivity between all pods before you configure Cloud Service Mesh. Additionally, if you join clusters that are not in the same project, they must be registered to the same fleet host project, and the clusters must be in a shared VPC configuration together on the same network. We also recommend that you have one project to host the shared VPC, and two service projects for creating clusters. For more information, see Setting up clusters with Shared VPC.

• If you use Istio CA, use the same custom root certificate for both clusters.

• If your Cloud Service Mesh is built on private clusters, we recommend creating a single subnet in the same VPC, otherwise, you must ensure that:

  1. The control planes can reach the remote private cluster control planes via the cluster private IPs.
  2. You can add the calling control planes' IP ranges to the remote private clusters' authorized networks. For more information, see Configure endpoint discovery between private clusters.

• The API server must be reachable by the other instances of the Cloud Service Mesh control plane in the multi-cluster mesh.

  • Ensure the clusters have global access enabled.
  • Ensure the Cloud Service Mesh control plane IP has been properly allowed via the allow list with the Master Authorized Network.

Setting project and cluster variables

1. Create the following environment variables for the project ID, cluster zone or region, cluster name, and context.

   export PROJECT_1=PROJECT_ID_1
   export LOCATION_1=CLUSTER_LOCATION_1
   export CLUSTER_1=CLUSTER_NAME_1
   export CTX_1="gke_${PROJECT_1}_${LOCATION_1}_${CLUSTER_1}"
   
   export PROJECT_2=PROJECT_ID_2
   export LOCATION_2=CLUSTER_LOCATION_2
   export CLUSTER_2=CLUSTER_NAME_2
   export CTX_2="gke_${PROJECT_2}_${LOCATION_2}_${CLUSTER_2}"
   

2. If these are newly created clusters, ensure to fetch credentials for each cluster with the following gcloud commands otherwise their associated context will not be available for use in the next steps of this guide.

   The commands depend on your cluster type, either regional or zonal:

   Regional

   gcloud container clusters get-credentials ${CLUSTER_1} --region ${LOCATION_1}
   gcloud container clusters get-credentials ${CLUSTER_2} --region ${LOCATION_2}
   

   Zonal

   gcloud container clusters get-credentials ${CLUSTER_1} --zone ${LOCATION_1}
   gcloud container clusters get-credentials ${CLUSTER_2} --zone ${LOCATION_2}
   

Create firewall rule

In some cases, you need to create a firewall rule to allow cross-cluster traffic. For example, you need to create a firewall rule if:

• You use different subnets for the clusters in your mesh.
• Your Pods open ports other than 443 and 15002.

GKE automatically adds firewall rules to each node to allow traffic within the same subnet. If your mesh contains multiple subnets, you must explicitly set up the firewall rules to allow cross-subnet traffic. You must add a new firewall rule for each subnet to allow the source IP CIDR blocks and targets ports of all the incoming traffic.

The following instructions allow communication between all clusters in your project or only between $CLUSTER_1 and $CLUSTER_2.

1. Gather information about your clusters' network.

   All project clusters

   If the clusters are in the same project, you can use the following command to allow communication between all clusters in your project. If there are clusters in your project that you don't want to expose, use the command in the Specific clusters tab.

   function join_by { local IFS="$1"; shift; echo "$*"; }
   ALL_CLUSTER_CIDRS=$(gcloud container clusters list --project $PROJECT_1 --format='value(clusterIpv4Cidr)' | sort | uniq)
   ALL_CLUSTER_CIDRS=$(join_by , $(echo "${ALL_CLUSTER_CIDRS}"))
   ALL_CLUSTER_NETTAGS=$(gcloud compute instances list --project $PROJECT_1 --format='value(tags.items.[0])' | sort | uniq)
   ALL_CLUSTER_NETTAGS=$(join_by , $(echo "${ALL_CLUSTER_NETTAGS}"))
   

   Specific clusters

   The following command allows communication between $CLUSTER_1 and $CLUSTER_2 and doesn't expose other clusters in your project.

   function join_by { local IFS="$1"; shift; echo "$*"; }
   ALL_CLUSTER_CIDRS=$(for P in $PROJECT_1 $PROJECT_2; do gcloud --project $P container clusters list --filter="name:($CLUSTER_1,$CLUSTER_2)" --format='value(clusterIpv4Cidr)'; done | sort | uniq)
   ALL_CLUSTER_CIDRS=$(join_by , $(echo "${ALL_CLUSTER_CIDRS}"))
   ALL_CLUSTER_NETTAGS=$(for P in $PROJECT_1 $PROJECT_2; do gcloud --project $P compute instances list  --filter="name:($CLUSTER_1,$CLUSTER_2)" --format='value(tags.items.[0])' ; done | sort | uniq)
   ALL_CLUSTER_NETTAGS=$(join_by , $(echo "${ALL_CLUSTER_NETTAGS}"))
   

2. Create the firewall rule.

   GKE

   gcloud compute firewall-rules create istio-multicluster-pods \
       --allow=tcp,udp,icmp,esp,ah,sctp \
       --direction=INGRESS \
       --priority=900 \
       --source-ranges="${ALL_CLUSTER_CIDRS}" \
       --target-tags="${ALL_CLUSTER_NETTAGS}" --quiet \
       --network=YOUR_NETWORK
   

   Autopilot

   TAGS=""
   for CLUSTER in ${CLUSTER_1} ${CLUSTER_2}
   do
       TAGS+=$(gcloud compute firewall-rules list --filter="Name:$CLUSTER*" --format="value(targetTags)" | uniq) && TAGS+=","
   done
   TAGS=${TAGS::-1}
   echo "Network tags for pod ranges are $TAGS"
   
   gcloud compute firewall-rules create asm-multicluster-pods \
       --allow=tcp,udp,icmp,esp,ah,sctp \
       --network=gke-cluster-vpc \
       --direction=INGRESS \
       --priority=900 --network=VPC_NAME \
       --source-ranges="${ALL_CLUSTER_CIDRS}" \
       --target-tags=$TAGS
   

Configure endpoint discovery

Note: For more information on endpoint discovery, refer to Endpoint discovery with multiple control planes.

The steps required to configure endpoint discovery depend on whether you prefer to use the declarative API across clusters in a fleet, or enable it manually on public clusters or private clusters.

Warning: If endpoint discovery is enabled between clusters, inter-cluster services are not able to communicate with each other without proper DNS configuration. This is because GKE services depend on each service's fully qualified domain name (FQDN) for traffic routing. You can enable DNS proxy, to allow inter-cluster discovery. DNS proxy captures all DNS requests and resolves them using information from the Control Plane.

Enable endpoint discovery between clusters with declarative API (preview)

Preview — Enable endpoint discovery

This feature is subject to the "Pre-GA Offerings Terms" in the General Service Terms section of the Service Specific Terms. Pre-GA features are available "as is" and might have limited support. For more information, see the launch stage descriptions.

You can enable endpoint discovery across clusters in a fleet by applying the config "multicluster_mode":"connected" in the in the asm-options configmap. Clusters with this config enabled in the same fleet will have cross-cluster service discovery automatically enabled between each other.

Warning: Do not use this method if your clusters already use existing in-cluster remote secrets.

This method is available for managed Cloud Service Mesh installations on all release channels. Before proceeding, you must have created a firewall rule.

For multiple projects, you must manually add FLEET_PROJECT_ID.svc.id.goog to trustDomainAliases in the revison's meshConfig if it's not already present.

Enable

If the asm-options configmap already exists in your cluster, then enable endpoint discovery for the cluster:

kubectl patch configmap/asm-options -n istio-system --type merge -p '{"data":{"multicluster_mode":"connected"}}'

If the asm-options configmap does not yet exist in your cluster, then create it with the associated data and enable endpoint discovery for the cluster:

kubectl --context ${CTX_1} create configmap asm-options -n istio-system --from-file <(echo '{"data":{"multicluster_mode":"connected"}}')

Disable

Disable endpoint discovery for a cluster:

kubectl patch configmap/asm-options -n istio-system --type merge -p '{"data":{"multicluster_mode":"manual"}}'

If you unregister a cluster from the fleet without disabling endpoint discovery, secrets could remain in the cluster. You must manually clean up any remaining secrets.

1. Run the following command to find secrets requiring cleanup:

   kubectl get secrets -n istio-system -l istio.io/owned-by=mesh.googleapis.com,istio/multiCluster=true
   

2. Delete each secret:

   kubectl delete secret SECRET_NAME
   

   Repeat this step for each remaining secret.

Configure endpoint discovery between public clusters

To configure endpoint discovery between GKE clusters, you run asmcli create-mesh. This command:

• Registers all clusters to the same fleet.
• Configures the mesh to trust the fleet workload identity.
• Creates remote secrets.

You can either specify the URI for each cluster or the path the kubeconfig file.

Cluster URI

In the following command, replace FLEET_PROJECT_ID with the project ID of the fleet host project and the cluster URI with the cluster name, zone or region, and project ID for each cluster. This example only shows two clusters, but you can run the command to enable endpoint discovery on additional clusters, subject to the maximum permitted number of clusters that you can add to your fleet.

./asmcli create-mesh \
    FLEET_PROJECT_ID \
    ${PROJECT_1}/${LOCATION_1}/${CLUSTER_1} \
    ${PROJECT_2}/${LOCATION_2}/${CLUSTER_2}

kubeconfig file

In the following command, replace FLEET_PROJECT_ID with the project ID of the fleet host project and PATH_TO_KUBECONFIG with the path to each kubeconfig file. This example only shows two clusters, but you can run the command to enable endpoint discovery on additional clusters, subject to the maximum permitted number of clusters that you can add to your fleet.

./asmcli create-mesh \
    FLEET_PROJECT_ID \
    PATH_TO_KUBECONFIG_1 \
    PATH_TO_KUBECONFIG_2

Configure endpoint discovery between private clusters

1. Configure remote secrets to allow API server access to the cluster to the other cluster's Cloud Service Mesh control plane. The commands depend on your Cloud Service Mesh type (either in-cluster or managed):

   Note: Ensure you run the following commands using the istioctl from the --output_dir location you specified when running asmcli install by executing istioctl commands in a way ../output_dir/istioctl.

   A. For in-cluster Cloud Service Mesh, you must configure the private IPs instead of public IPs because the public IPs are not accessible:

   PRIV_IP=`gcloud container clusters describe "${CLUSTER_1}" --project "${PROJECT_1}" \
    --zone "${LOCATION_1}" --format "value(privateClusterConfig.privateEndpoint)"`
   
   ./istioctl x create-remote-secret --context=${CTX_1} --name=${CLUSTER_1} --server=https://${PRIV_IP} > ${CTX_1}.secret
   

   PRIV_IP=`gcloud container clusters describe "${CLUSTER_2}" --project "${PROJECT_2}" \
    --zone "${LOCATION_2}" --format "value(privateClusterConfig.privateEndpoint)"`
   
   ./istioctl x create-remote-secret --context=${CTX_2} --name=${CLUSTER_2} --server=https://${PRIV_IP} > ${CTX_2}.secret
   

   B. For Managed Cloud Service Mesh:

   PUBLIC_IP=`gcloud container clusters describe "${CLUSTER_1}" --project "${PROJECT_1}" \
    --zone "${LOCATION_1}" --format "value(privateClusterConfig.publicEndpoint)"`
   
   ./istioctl x create-remote-secret --context=${CTX_1} --name=${CLUSTER_1} --server=https://${PUBLIC_IP} > ${CTX_1}.secret
   

   PUBLIC_IP=`gcloud container clusters describe "${CLUSTER_2}" --project "${PROJECT_2}" \
    --zone "${LOCATION_2}" --format "value(privateClusterConfig.publicEndpoint)"`
   
   ./istioctl x create-remote-secret --context=${CTX_2} --name=${CLUSTER_2} --server=https://${PUBLIC_IP} > ${CTX_2}.secret
   

2. Apply the new secrets into the clusters:

   kubectl apply -f ${CTX_1}.secret --context=${CTX_2}
   

   kubectl apply -f ${CTX_2}.secret --context=${CTX_1}
   

Configure authorized networks for private clusters

Follow this section only if all of the following conditions apply to your mesh:

• You are using private clusters.
• The clusters do not belong to the same subnet.
• The clusters have enabled authorized networks.

When deploying multiple private clusters, the Cloud Service Mesh control plane in each cluster needs to call the GKE control plane of the remote clusters. To allow traffic, you need to add the Pod address range in the calling cluster to the authorized networks of the remote clusters.

1. Get the Pod IP CIDR block for each cluster:

   POD_IP_CIDR_1=`gcloud container clusters describe ${CLUSTER_1} --project ${PROJECT_1} --zone ${LOCATION_1} \
     --format "value(ipAllocationPolicy.clusterIpv4CidrBlock)"`
   

   POD_IP_CIDR_2=`gcloud container clusters describe ${CLUSTER_2} --project ${PROJECT_2} --zone ${LOCATION_2} \
     --format "value(ipAllocationPolicy.clusterIpv4CidrBlock)"`
   

2. Add the Kubernetes cluster Pod IP CIDR blocks to the remote clusters:

   EXISTING_CIDR_1=`gcloud container clusters describe ${CLUSTER_1} --project ${PROJECT_1} --zone ${LOCATION_1} \
    --format "value(masterAuthorizedNetworksConfig.cidrBlocks.cidrBlock)"`
   gcloud container clusters update ${CLUSTER_1} --project ${PROJECT_1} --zone ${LOCATION_1} \
   --enable-master-authorized-networks \
   --master-authorized-networks ${POD_IP_CIDR_2},${EXISTING_CIDR_1//;/,}
   

   EXISTING_CIDR_2=`gcloud container clusters describe ${CLUSTER_2} --project ${PROJECT_2} --zone ${LOCATION_2} \
    --format "value(masterAuthorizedNetworksConfig.cidrBlocks.cidrBlock)"`
   gcloud container clusters update ${CLUSTER_2} --project ${PROJECT_2} --zone ${LOCATION_2} \
   --enable-master-authorized-networks \
   --master-authorized-networks ${POD_IP_CIDR_1},${EXISTING_CIDR_2//;/,}
   

   For more information, see Creating a cluster with authorized networks.

3. Verify that the authorized networks are updated:

   gcloud container clusters describe ${CLUSTER_1} --project ${PROJECT_1} --zone ${LOCATION_1} \
    --format "value(masterAuthorizedNetworksConfig.cidrBlocks.cidrBlock)"
   

   gcloud container clusters describe ${CLUSTER_2} --project ${PROJECT_2} --zone ${LOCATION_2} \
    --format "value(masterAuthorizedNetworksConfig.cidrBlocks.cidrBlock)"
   

Enable control plane global access

Follow this section only if all of the following conditions apply to your mesh:

• You are using private clusters.
• You use different regions for the clusters in your mesh.

You must enable control plane global access to allow Cloud Service Mesh control plane in each cluster to call the GKE control plane of the remote clusters.

1. Enable control plane global access:

   gcloud container clusters update ${CLUSTER_1} --project ${PROJECT_1} --zone ${LOCATION_1} \
    --enable-master-global-access
   

   gcloud container clusters update ${CLUSTER_2} --project ${PROJECT_2} --zone ${LOCATION_2} \
    --enable-master-global-access
   

2. Verify that control plane global access in enabled:

   gcloud container clusters describe ${CLUSTER_1} --project ${PROJECT_1} --zone ${LOCATION_1}
   

   gcloud container clusters describe ${CLUSTER_2} --project ${PROJECT_2} --zone ${LOCATION_2}
   

   The privateClusterConfig section in the output displays the status of masterGlobalAccessConfig.

Verify multicluster connectivity

Platform note: These steps are only applicable if you are using an in-cluster control plane. For managed Cloud Service Mesh, see Injecting sidecar proxies or Apply the managed data plane.

This section explains how to deploy the sample HelloWorld and Sleep services to your multi-cluster environment to verify that cross-cluster load balancing works.

Note: These sample services are located in the Istio samples directory included in the istioctl tarball (located in OUTPUT_DIR/istio-${ASM_VERSION%+*}/samples) and not the samples directory downloaded by asmcli (located in OUTPUT_DIR/samples).

Set variable for samples directory

1. Navigate to where asmcli was downloaded, and run the following command to set ASM_VERSION

   export ASM_VERSION="$(./asmcli --version)"
   

2. Set a working folder to the samples that you use to verify that cross-cluster load balancing works. The samples are located in a subdirectory in the --output_dir directory that you specified in the asmcli install command. In the following command, change OUTPUT_DIR to the directory that you specified in --output_dir.

   export SAMPLES_DIR=OUTPUT_DIR/istio-${ASM_VERSION%+*}
   

Enable sidecar injection

1. Create the sample namespace in each cluster.

   for CTX in ${CTX_1} ${CTX_2}
   do
       kubectl create --context=${CTX} namespace sample
   done
   

2. Enable the namespace for injection. The steps depend on your control plane implementation.

   Managed (TD)

   1. Apply the default injection label to the namespace:

   for CTX in ${CTX_1} ${CTX_2}
   do
      kubectl label --context=${CTX} namespace sample \
         istio.io/rev- istio-injection=enabled --overwrite
   done
   

   Managed (Istiod)

   Recommended: Run the following command to apply the default injection label to the namespace:

    for CTX in ${CTX_1} ${CTX_2}
    do
       kubectl label --context=${CTX} namespace sample \
          istio.io/rev- istio-injection=enabled --overwrite
    done
   

   If you are an existing user with the Managed Istiod control plane: We recommend that you use default injection, but revision-based injection is supported. Use the following instructions:

   1. Run the following command to locate the available release channels:

      kubectl -n istio-system get controlplanerevision
      

      The output is similar to the following:

      NAME                AGE
      asm-managed-rapid   6d7h
      

      Note: If two control plane revisions appear in the earlier list, remove one. Having multiple control plane channels in the cluster is not supported.

      In the output, the value under the NAME column is the revision label that corresponds to the available release channel for the Cloud Service Mesh version.

   2. Apply the revision label to the namespace:

      for CTX in ${CTX_1} ${CTX_2}
      do
        kubectl label --context=${CTX} namespace sample \
           istio-injection- istio.io/rev=REVISION_LABEL --overwrite
      done
      

Install the HelloWorld service

Note: The HelloWorld service example uses Docker Hub. In a private cluster, the container runtime can pull container images from Artifact Registry by default. The container runtime cannot pull images from any other container image registry on the internet. You can download the image and push to Artifact Registry or use Cloud NAT to provide outbound internet access for certain private nodes. For more information, see Migrate external containers and Creating a private cluster.

• Create the HelloWorld service in both clusters:

  kubectl create --context=${CTX_1} \
      -f ${SAMPLES_DIR}/samples/helloworld/helloworld.yaml \
      -l service=helloworld -n sample
  

  kubectl create --context=${CTX_2} \
      -f ${SAMPLES_DIR}/samples/helloworld/helloworld.yaml \
      -l service=helloworld -n sample
  

Deploy HelloWorld v1 and v2 to each cluster

1. Deploy HelloWorld v1 to CLUSTER_1 and v2 to CLUSTER_2, which helps later to verify cross-cluster load balancing:

   kubectl create --context=${CTX_1} \
     -f ${SAMPLES_DIR}/samples/helloworld/helloworld.yaml \
     -l version=v1 -n sample

   kubectl create --context=${CTX_2} \
     -f ${SAMPLES_DIR}/samples/helloworld/helloworld.yaml \
     -l version=v2 -n sample

2. Confirm HelloWorld v1 and v2 are running using the following commands. Verify that the output is similar to that shown.:

   kubectl get pod --context=${CTX_1} -n sample

   NAME                            READY     STATUS    RESTARTS   AGE
   helloworld-v1-86f77cd7bd-cpxhv  2/2       Running   0          40s

   kubectl get pod --context=${CTX_2} -n sample

   NAME                            READY     STATUS    RESTARTS   AGE
   helloworld-v2-758dd55874-6x4t8  2/2       Running   0          40s

Deploy the Sleep service

1. Deploy the Sleep service to both clusters. This pod generates artificial network traffic for demonstration purposes:

   for CTX in ${CTX_1} ${CTX_2}
   do
       kubectl apply --context=${CTX} \
           -f ${SAMPLES_DIR}/samples/sleep/sleep.yaml -n sample
   done
   

2. Wait for the Sleep service to start in each cluster. Verify that the output is similar to that shown:

   kubectl get pod --context=${CTX_1} -n sample -l app=sleep

   NAME                             READY   STATUS    RESTARTS   AGE
   sleep-754684654f-n6bzf           2/2     Running   0          5s

   kubectl get pod --context=${CTX_2} -n sample -l app=sleep

   NAME                             READY   STATUS    RESTARTS   AGE
   sleep-754684654f-dzl9j           2/2     Running   0          5s

Verify cross-cluster load balancing

Call the HelloWorld service several times and check the output to verify alternating replies from v1 and v2:

1. Call the HelloWorld service:

   kubectl exec --context="${CTX_1}" -n sample -c sleep \
       "$(kubectl get pod --context="${CTX_1}" -n sample -l \
       app=sleep -o jsonpath='{.items[0].metadata.name}')" \
       -- /bin/sh -c 'for i in $(seq 1 20); do curl -sS helloworld.sample:5000/hello; done'
   

   The output is similar to that shown:

   Hello version: v2, instance: helloworld-v2-758dd55874-6x4t8
   Hello version: v1, instance: helloworld-v1-86f77cd7bd-cpxhv
   ...

2. Call the HelloWorld service again:

   kubectl exec --context="${CTX_2}" -n sample -c sleep \
       "$(kubectl get pod --context="${CTX_2}" -n sample -l \
       app=sleep -o jsonpath='{.items[0].metadata.name}')" \
       -- /bin/sh -c 'for i in $(seq 1 20); do curl -sS helloworld.sample:5000/hello; done'
   

   The output is similar to that shown:

   Hello version: v2, instance: helloworld-v2-758dd55874-6x4t8
   Hello version: v1, instance: helloworld-v1-86f77cd7bd-cpxhv
   ...

Congratulations, you've verified your load-balanced, multi-cluster Cloud Service Mesh!

Clean up HelloWorld service

When you finish verifying load balancing, remove the HelloWorld and Sleep service from your cluster.

kubectl delete ns sample --context ${CTX_1}
kubectl delete ns sample --context ${CTX_2}