Install Multiple Istio Control Planes in a Single Cluster

This guide walks you through the process of installing multiple Istio control planes within a single cluster and then a way to scope workloads to specific control planes. This deployment model has a single Kubernetes control plane with multiple Istio control planes and meshes. The separation between the meshes is provided by Kubernetes namespaces and RBAC.

Multiple meshes in a single cluster
Multiple meshes in a single cluster

Using discoverySelectors, you can scope Kubernetes resources in a cluster to specific namespaces managed by an Istio control plane. This includes the Istio custom resources (e.g., Gateway, VirtualService, DestinationRule, etc.) used to configure the mesh. Furthermore, discoverySelectors can be used to configure which namespaces should include the istio-ca-root-cert config map for a particular Istio control plane. Together, these functions allow mesh operators to specify the namespaces for a given control plane, enabling soft multi-tenancy for multiple meshes based on the boundary of one or more namespaces. This guide uses discoverySelectors, along with the revisions capability of Istio, to demonstrate how two meshes can be deployed on a single cluster, each working with a properly scoped subset of the cluster’s resources.

Before you begin

This guide requires that you have a Kubernetes cluster with any of the supported Kubernetes versions: 1.24, 1.25, 1.26, 1.27.

This cluster will host two control planes installed in two different system namespaces. The mesh application workloads will run in multiple application-specific namespaces, each namespace associated with one or the other control plane based on revision and discovery selector configurations.

Cluster configuration

Deploying multiple control planes

Deploying multiple Istio control planes on a single cluster can be achieved by using different system namespaces for each control plane. Istio revisions and discoverySelectors are then used to scope the resources and workloads that are managed by each control plane.

  1. Create the first system namespace, usergroup-1, and deploy istiod in it:

    $ kubectl create ns usergroup-1
    $ kubectl label ns usergroup-1 usergroup=usergroup-1
    $ istioctl install -y -f - <<EOF
    apiVersion: install.istio.io/v1alpha1
    kind: IstioOperator
    metadata:
      namespace: usergroup-1
    spec:
      profile: minimal
      revision: usergroup-1
      meshConfig:
        discoverySelectors:
          - matchLabels:
              usergroup: usergroup-1
      values:
        global:
          istioNamespace: usergroup-1
        pilot:
          env:
            ENABLE_ENHANCED_RESOURCE_SCOPING: true
    EOF
    
  2. Create the second system namespace, usergroup-2, and deploy istiod in it:

    $ kubectl create ns usergroup-2
    $ kubectl label ns usergroup-2 usergroup=usergroup-2
    $ istioctl install -y -f - <<EOF
    apiVersion: install.istio.io/v1alpha1
    kind: IstioOperator
    metadata:
      namespace: usergroup-2
    spec:
      profile: minimal
      revision: usergroup-2
      meshConfig:
        discoverySelectors:
          - matchLabels:
              usergroup: usergroup-2
      values:
        global:
          istioNamespace: usergroup-2
        pilot:
          env:
            ENABLE_ENHANCED_RESOURCE_SCOPING: true
    EOF
    
  3. Deploy a policy for workloads in the usergroup-1 namespace to only accept mutual TLS traffic:

    $ kubectl apply -f - <<EOF
    apiVersion: security.istio.io/v1beta1
    kind: PeerAuthentication
    metadata:
      name: "usergroup-1-peerauth"
      namespace: "usergroup-1"
    spec:
      mtls:
        mode: STRICT
    EOF
    
  4. Deploy a policy for workloads in the usergroup-2 namespace to only accept mutual TLS traffic:

    $ kubectl apply -f - <<EOF
    apiVersion: security.istio.io/v1beta1
    kind: PeerAuthentication
    metadata:
      name: "usergroup-2-peerauth"
      namespace: "usergroup-2"
    spec:
      mtls:
        mode: STRICT
    EOF
    

Verify the multiple control plane creation

  1. Check the labels on the system namespaces for each control plane:

    $ kubectl get ns usergroup-1 usergroup2 --show-labels
    NAME              STATUS   AGE     LABELS
    usergroup-1       Active   13m     kubernetes.io/metadata.name=usergroup-1,usergroup=usergroup-1
    usergroup-2       Active   12m     kubernetes.io/metadata.name=usergroup-2,usergroup=usergroup-2
    
  2. Verify the control planes are deployed and running:

    $ kubectl get pods -n usergroup-1
    NAMESPACE     NAME                                     READY   STATUS    RESTARTS         AGE
    usergroup-1   istiod-usergroup-1-5ccc849b5f-wnqd6      1/1     Running   0                12m
    
    $ kubectl get pods -n usergroup-2
    NAMESPACE     NAME                                     READY   STATUS    RESTARTS         AGE
    usergroup-2   istiod-usergroup-2-658d6458f7-slpd9      1/1     Running   0                12m
    

    You will notice that one istiod deployment per usergroup is created in the specified namespaces.

  3. Run the following commands to list the installed webhooks:

    $ kubectl get validatingwebhookconfiguration
    NAME                                      WEBHOOKS   AGE
    istio-validator-usergroup-1-usergroup-1   1          18m
    istio-validator-usergroup-2-usergroup-2   1          18m
    istiod-default-validator                  1          18m
    
    $ kubectl get mutatingwebhookconfiguration
    NAME                                             WEBHOOKS   AGE
    istio-revision-tag-default-usergroup-1           4          18m
    istio-sidecar-injector-usergroup-1-usergroup-1   2          19m
    istio-sidecar-injector-usergroup-2-usergroup-2   2          18m
    

    Note that the output includes istiod-default-validator and istio-revision-tag-default-usergroup-1, which are the default webhook configurations used for handling requests coming from resources which are not associated with any revision. In a fully scoped environment where every control plane is associated with its resources through proper namespace labeling, there is no need for these default webhook configurations. They should never be invoked.

Deploy application workloads per usergroup

  1. Create three application namespaces:

    $ kubectl create ns app-ns-1
    $ kubectl create ns app-ns-2
    $ kubectl create ns app-ns-3
    
  2. Label each namespace to associate them with their respective control planes:

    $ kubectl label ns app-ns-1 usergroup=usergroup-1 istio.io/rev=usergroup-1
    $ kubectl label ns app-ns-2 usergroup=usergroup-2 istio.io/rev=usergroup-2
    $ kubectl label ns app-ns-3 usergroup=usergroup-2 istio.io/rev=usergroup-2
    
  3. Deploy one sleep and httpbin application per namespace:

    $ kubectl -n app-ns-1 apply -f samples/sleep/sleep.yaml
    $ kubectl -n app-ns-1 apply -f samples/httpbin/httpbin.yaml
    $ kubectl -n app-ns-2 apply -f samples/sleep/sleep.yaml
    $ kubectl -n app-ns-2 apply -f samples/httpbin/httpbin.yaml
    $ kubectl -n app-ns-3 apply -f samples/sleep/sleep.yaml
    $ kubectl -n app-ns-3 apply -f samples/httpbin/httpbin.yaml
    
  4. Wait a few seconds for the httpbin and sleep pods to be running with sidecars injected:

    $ kubectl get pods -n app-ns-1
    NAME                      READY   STATUS    RESTARTS   AGE
    httpbin-9dbd644c7-zc2v4   2/2     Running   0          115m
    sleep-78ff5975c6-fml7c    2/2     Running   0          115m
    
    $ kubectl get pods -n app-ns-2
    NAME                      READY   STATUS    RESTARTS   AGE
    httpbin-9dbd644c7-sd9ln   2/2     Running   0          115m
    sleep-78ff5975c6-sz728    2/2     Running   0          115m
    
    $ kubectl get pods -n app-ns-3
    NAME                      READY   STATUS    RESTARTS   AGE
    httpbin-9dbd644c7-8ll27   2/2     Running   0          115m
    sleep-78ff5975c6-sg4tq    2/2     Running   0          115m
    

Verify the application to control plane mapping

Now that the applications are deployed, you can use the istioctl ps command to confirm that the application workloads are managed by their respective control plane, i.e., app-ns-1 is managed by usergroup-1, app-ns-2 and app-ns-3 are managed by usergroup-2:

$ istioctl ps -i usergroup-1
NAME                                 CLUSTER        CDS        LDS        EDS        RDS          ECDS         ISTIOD                                  VERSION
httpbin-9dbd644c7-hccpf.app-ns-1     Kubernetes     SYNCED     SYNCED     SYNCED     SYNCED       NOT SENT     istiod-usergroup-1-5ccc849b5f-wnqd6     1.17-alpha.f5212a6f7df61fd8156f3585154bed2f003c4117
sleep-78ff5975c6-9zb77.app-ns-1      Kubernetes     SYNCED     SYNCED     SYNCED     SYNCED       NOT SENT     istiod-usergroup-1-5ccc849b5f-wnqd6     1.17-alpha.f5212a6f7df61fd8156f3585154bed2f003c4117
$ istioctl ps -i usergroup-2
NAME                                 CLUSTER        CDS        LDS        EDS        RDS          ECDS         ISTIOD                                  VERSION
httpbin-9dbd644c7-vvcqj.app-ns-3     Kubernetes     SYNCED     SYNCED     SYNCED     SYNCED       NOT SENT     istiod-usergroup-2-658d6458f7-slpd9     1.17-alpha.f5212a6f7df61fd8156f3585154bed2f003c4117
httpbin-9dbd644c7-xzgfm.app-ns-2     Kubernetes     SYNCED     SYNCED     SYNCED     SYNCED       NOT SENT     istiod-usergroup-2-658d6458f7-slpd9     1.17-alpha.f5212a6f7df61fd8156f3585154bed2f003c4117
sleep-78ff5975c6-fthmt.app-ns-2      Kubernetes     SYNCED     SYNCED     SYNCED     SYNCED       NOT SENT     istiod-usergroup-2-658d6458f7-slpd9     1.17-alpha.f5212a6f7df61fd8156f3585154bed2f003c4117
sleep-78ff5975c6-nxtth.app-ns-3      Kubernetes     SYNCED     SYNCED     SYNCED     SYNCED       NOT SENT     istiod-usergroup-2-658d6458f7-slpd9     1.17-alpha.f5212a6f7df61fd8156f3585154bed2f003c4117

Verify the application connectivity is ONLY within the respective usergroup

  1. Send a request from the sleep pod in app-ns-1 in usergroup-1 to the httpbin service in app-ns-2 in usergroup-2. The communication should fail:

    $ kubectl -n app-ns-1 exec "$(kubectl -n app-ns-1 get pod -l app=sleep -o jsonpath={.items..metadata.name})" -c sleep -- curl -sIL http://httpbin.app-ns-2.svc.cluster.local:8000
    HTTP/1.1 503 Service Unavailable
    content-length: 95
    content-type: text/plain
    date: Sat, 24 Dec 2022 06:54:54 GMT
    server: envoy
    
  2. Send a request from the sleep pod in app-ns-2 in usergroup-2 to the httpbin service in app-ns-3 in usergroup-2. The communication should work:

    $ kubectl -n app-ns-2 exec "$(kubectl -n app-ns-2 get pod -l app=sleep -o jsonpath={.items..metadata.name})" -c sleep -- curl -sIL http://httpbin.app-ns-3.svc.cluster.local:8000
    HTTP/1.1 200 OK
    server: envoy
    date: Thu, 22 Dec 2022 15:01:36 GMT
    content-type: text/html; charset=utf-8
    content-length: 9593
    access-control-allow-origin: *
    access-control-allow-credentials: true
    x-envoy-upstream-service-time: 3
    

Cleanup

  1. Clean up the first usergroup:

    $ istioctl uninstall --revision usergroup-1
    $ kubectl delete ns app-ns-1 usergroup-1
    
  2. Clean up the second usergroup:

    $ istioctl uninstall --revision usergroup-2
    $ kubectl delete ns app-ns-2 app-ns-3 usergroup-2
    
Was this information useful?
Do you have any suggestions for improvement?

Thanks for your feedback!