Skip to content

KubeStellar Usage Examples#

This document shows some simple examples of using the release that contains this version of this document. See also the doc about using an existing hosting cluster and multi-machine scenarios for considerations for different scenarios. For historical examples, see our blog posts.

Each example scenario concludes with instructions on how to undo its effects, restoring the state established by the common setup so that it is easy to run through several scenarios without having to repeat the common setup.

There are also end-to-end (E2E) tests that are based on scenario 4 and an extended variant of scenario 1. These tests normally exercise the copy of the repo containing them (rather than a release). They can alternatively test a release. See the e2e tests (in test/e2e). Contributors can run these tests, and CI includes checking that these E2E tests pass. These tests are written in bash, so that contributors can easily follow them.

Prereqs#

See pre-reqs.

Common Setup#

The following steps establish an initial state used in the examples below.

There are to main way to proceed:

  1. Step-by-step manual setup
  2. Using KubeStellar Core Helm chart

Create and register two clusters with OCM#

Follow the steps to create and register two clusters with OCM.

Scenario 1 - multi-cluster workload deployment with kubectl#

This scenario proceeds from the state established by the common setup.

Check for available clusters with label location-group=edge

kubectl --context its1 get managedclusters -l location-group=edge

Create a BindingPolicy to deliver an app to all clusters in wds1:

kubectl --context wds1 apply -f - <<EOF
apiVersion: control.kubestellar.io/v1alpha1
kind: BindingPolicy
metadata:
  name: nginx-bpolicy
spec:
  clusterSelectors:
  - matchLabels: {"location-group":"edge"}
  downsync:
  - objectSelectors:
    - matchLabels: {"app.kubernetes.io/name":"nginx"}
EOF

This BindingPolicy configuration determines where to deploy the workload by using the label selector expressions found in clusterSelectors. It also specifies what to deploy through the downsync.labelSelectors expressions. Each matchLabels expression is a criterion for selecting a set of objects based on their labels. Other criteria can be added to filter objects based on their namespace, api group, resource, and name. If these criteria are not specified, all objects with the matching labels are selected. If an object has multiple labels, it is selected only if it matches all the labels in the matchLabels expression. If there are multiple objectSelectors, an object is selected if it matches any of them.

Now deploy the app:

kubectl --context wds1 apply -f - <<EOF
apiVersion: v1
kind: Namespace
metadata:
  labels:
    app.kubernetes.io/name: nginx
  name: nginx
---
apiVersion: apps/v1
kind: Deployment
metadata:
  name: nginx-deployment
  namespace: nginx
  labels:
    app.kubernetes.io/name: nginx
spec:
  replicas: 1
  selector:
    matchLabels:
      app: nginx
  template:
    metadata:
      labels:
        app: nginx
    spec:
      containers:
      - name: nginx
        image: public.ecr.aws/nginx/nginx:latest
        ports:
        - containerPort: 80
EOF

Verify that manifestworks wrapping the objects have been created in the mailbox namespaces:

kubectl --context its1 get manifestworks -n cluster1
kubectl --context its1 get manifestworks -n cluster2

Verify that the deployment has been created in both clusters

kubectl --context cluster1 get deployments -n nginx
kubectl --context cluster2 get deployments -n nginx

Please note, in line with Kubernetes’ best practices, the order in which you apply a BindingPolicy and the objects doesn’t affect the outcome. You can apply the BindingPolicy first followed by the objects, or vice versa. The result remains consistent because the binding controller identifies any changes in either the BindingPolicy or the objects, triggering the start of the reconciliation loop.

[Optional] Teardown Scenario 1#

kubectl --context wds1 delete ns nginx
kubectl --context wds1 delete bindingpolicies nginx-bpolicy

Scenario 2 - using the hosting cluster as WDS to deploy a custom resource#

This scenario follows on from the state established by scenario 1.

The hosting cluster can act as a Workload Description Space (WDS) to distribute your workloads to multiple clusters. This feature works well for Custom Resources, but not for standard Kubernetes resources (deployments, pods, replicasets, etc.). The reason is that the hosting cluster’s controller manager creates pods for those resources on the hosting cluster itself, while the Kubestellar controller copies them to the Workload Execution Clusters (WECs). You can use any Custom Resource to wrap any Kubernetes object you want to send to the WECs. But if you have operators or controllers on the hosting cluster that work on the Custom Resource you want to send, make sure they don’t create workloads on the hosting cluster that you did not intend to create there.

In order to run this scenario using the post-create-hook method you need the raise the permissions for the kubeflex controller manager:

kubectl --context kind-kubeflex apply -f - <<EOF
apiVersion: rbac.authorization.k8s.io/v1
kind: ClusterRoleBinding
metadata:
  name: kubeflex-manager-cluster-admin-rolebinding
roleRef:
  apiGroup: rbac.authorization.k8s.io
  kind: ClusterRole
  name: cluster-admin
subjects:
- kind: ServiceAccount
  name: kubeflex-controller-manager
  namespace: kubeflex-system
EOF

To create a second WDS based on the hosting cluster, run the commands:

kflex create wds2 -t host

helm --kube-context kind-kubeflex upgrade --install ocm-transport-plugin oci://ghcr.io/kubestellar/ocm-transport-plugin/chart/ocm-transport-plugin --version ${OCM_TRANSPORT_PLUGIN} \
    --set transport_cp_name=its1 \
    --set wds_cp_name=wds2 \
    -n wds2-system

where the -t host option specifies a control plane of type host. You can only create one control plane of type host.

In this example, we use the helm chart method to install the KubeStellar controller manager for the hosting cluster so that we can pass additional startup options.

Label the wds2 control plane as type wds:

kubectl label cp wds2 kflex.kubestellar.io/cptype=wds

For this example, we use the AppWrapper custom resource defined in the multi cluster app dispatcher project.

Install the AppWrapper CRD in the WDS and the WECs.

clusters=(wds2 cluster1 cluster2);
  for cluster in "${clusters[@]}"; do
  kubectl --context ${cluster} apply -f https://raw.githubusercontent.com/project-codeflare/multi-cluster-app-dispatcher/v1.39.0/config/crd/bases/workload.codeflare.dev_appwrappers.yaml
done

Apply the kubestellar controller-manager helm chart with the option to allow only delivery of objects with api group workload.codeflare.dev

helm --kube-context kind-kubeflex upgrade --install -n wds2-system kubestellar oci://ghcr.io/kubestellar/kubestellar/controller-manager-chart --version ${KUBESTELLAR_VERSION} --set ControlPlaneName=wds2 --set APIGroups=workload.codeflare.dev

Check that the kubestellar controller for wds2 is started:

kubectl get deployments.apps -n wds2-system kubestellar-controller-manager

If desired, you may remove the kubeflex-manager-cluster-admin-rolebinding after the kubestellar-controller-manager is started, with the command kubectl --context kind-kubeflex delete clusterrolebinding kubeflex-manager-cluster-admin-rolebinding

Run the following comamand to give permission for the Klusterlet to operate on the appwrapper cluster resource.

clusters=(cluster1 cluster2);
for cluster in "${clusters[@]}"; do
kubectl --context ${cluster} apply -f - <<EOF
apiVersion: rbac.authorization.k8s.io/v1
kind: ClusterRole
metadata:
  name: appwrappers-access
rules:
- apiGroups: ["workload.codeflare.dev"]
  resources: ["appwrappers"]
  verbs: ["get", "list", "watch", "create", "update", "patch"]
---
apiVersion: rbac.authorization.k8s.io/v1
kind: ClusterRoleBinding
metadata:
  name: klusterlet-appwrappers-access
roleRef:
  apiGroup: rbac.authorization.k8s.io
  kind: ClusterRole
  name: appwrappers-access
subjects:
- kind: ServiceAccount
  name: klusterlet-work-sa
  namespace: open-cluster-management-agent
EOF
done

This step will be eventually automated, see this issue for more details.

Next, apply an appwrapper object to wds2:

kubectl --context wds2 apply -f  https://raw.githubusercontent.com/project-codeflare/multi-cluster-app-dispatcher/v1.39.0/test/yaml/0008-aw-default.yaml

Label the appwrapper to match the binding policy:

kubectl --context wds2 label appwrappers.workload.codeflare.dev defaultaw-schd-spec-with-timeout-1 app.kubernetes.io/part-of=my-appwrapper-app

Finally, apply the BindingPolicy:

kubectl --context wds2 apply -f - <<EOF
apiVersion: control.kubestellar.io/v1alpha1
kind: BindingPolicy
metadata:
  name: aw-bpolicy
spec:
  clusterSelectors:
  - matchLabels: {"location-group":"edge"}
  downsync:
  - objectSelectors:
    - matchLabels: {"app.kubernetes.io/part-of":"my-appwrapper-app"}
EOF

Check that the app wrapper has been delivered to both clusters:

kubectl --context cluster1 get appwrappers
kubectl --context cluster2 get appwrappers

[Optional] Teardown Scenario 2#

kubectl --context wds2 delete bindingpolicies aw-bpolicy
kubectl --context wds2 delete appwrappers --all

Wait until the following commands show no appwrappers in cluster1 and cluster2.

kubectl --context cluster1 get appwrappers -A
kubectl --context cluster2 get appwrappers -A

Then continue.

for cluster in cluster1 cluster2; do
  kubectl --context $cluster delete clusterroles appwrappers-access
  kubectl --context $cluster delete clusterrolebindings klusterlet-appwrappers-access
done

If you have not already done so, then do the following command.

kubectl --context kind-kubeflex delete clusterrolebinding kubeflex-manager-cluster-admin-rolebinding

Continue as follows.

helm --kube-context kind-kubeflex uninstall -n wds2-system kubestellar

clusters=(wds2 cluster1 cluster2);
  for cluster in "${clusters[@]}"; do
  kubectl --context ${cluster} delete -f https://raw.githubusercontent.com/project-codeflare/multi-cluster-app-dispatcher/v1.39.0/config/crd/bases/workload.codeflare.dev_appwrappers.yaml
done

kflex delete wds2

Scenario 3 - multi-cluster workload deployment with helm#

This scenario proceeds from the state established by the common setup.

Create a BindingPolicy for the helm chart app:

kubectl --context wds1 apply -f - <<EOF
apiVersion: control.kubestellar.io/v1alpha1
kind: BindingPolicy
metadata:
  name: postgres-bpolicy
spec:
  clusterSelectors:
  - matchLabels: {"location-group":"edge"}
  downsync:
  - objectSelectors:
    - matchLabels: {
      "app.kubernetes.io/managed-by": Helm,
      "app.kubernetes.io/instance": postgres}
EOF

Note that helm sets app.kubernetes.io/instance to the name of the installed release.

Create and label the namespace and install the chart:

kubectl --context wds1 create ns postgres-system
kubectl --context wds1 label ns postgres-system app.kubernetes.io/managed-by=Helm app.kubernetes.io/instance=postgres
helm --kube-context wds1 install -n postgres-system postgres oci://registry-1.docker.io/bitnamicharts/postgresql

Verify that statefulset has been created in both clusters

kubectl --context cluster1 get statefulsets -n postgres-system
kubectl --context cluster2 get statefulsets -n postgres-system

[Optional] Propagate helm metadata Secret to managed clusters#

Run "helm list" on the wds1:

$ helm --kube-context wds1 list -n postgres-system
NAME            NAMESPACE       REVISION        UPDATED                                 STATUS       CHART                    APP VERSION
postgres        postgres-system 1               2023-10-31 13:39:52.550071 -0400 EDT    deployed     postgresql-13.2.0        16.0.0

And try that on the managed clusters

$ helm list --kube-context cluster1 -n postgres-system
: returns empty
$ helm list --kube-context cluster2 -n postgres-system
: returns empty

This is because Helm creates a Secret object to hold its metadata about a "release" (chart instance) but Helm does not apply the usual labels to that object, so it is not selected by the BindingPolicy above and thus does not get delivered. The workload is functioning in the WECs, but helm list does not recognize its handiwork there. That labeling could be done for example with:

kubectl --context wds1 label secret -n postgres-system $(kubectl --context wds1 get secrets -n postgres-system -l name=postgres -l owner=helm  -o jsonpath='{.items[0].metadata.name}') app.kubernetes.io/managed-by=Helm app.kubernetes.io/instance=postgres

Verify that the chart shows up on the managed clusters:

helm list --kube-context cluster1 -n postgres-system
helm list --kube-context cluster2 -n postgres-system

Implementing this in a controller for automated propagation of helm metadata is tracked in this issue.

[Optional] Teardown Scenario 3#

helm --kube-context wds1 uninstall -n postgres-system postgres
kubectl --context wds1 delete ns postgres-system
kubectl --context wds1 delete bindingpolicies postgres-bpolicy

Scenario 4 - Singleton status#

This scenario proceeds from the state established by the common setup.

This scenario shows how to get the full status updated when setting wantSingletonReportedState in the BindingPolicy. This still an experimental feature.

Apply a BindingPolicy with the wantSingletonReportedState flag set:

kubectl --context wds1 apply -f - <<EOF
apiVersion: control.kubestellar.io/v1alpha1
kind: BindingPolicy
metadata:
  name: nginx-singleton-bpolicy
spec:
  wantSingletonReportedState: true
  clusterSelectors:
  - matchLabels: {"name":"cluster1"}
  downsync:
  - objectSelectors:
    - matchLabels: {"app.kubernetes.io/name":"nginx-singleton"}
EOF

Apply a new deployment for the singleton BindingPolicy:

kubectl --context wds1 apply -f - <<EOF
apiVersion: apps/v1
kind: Deployment
metadata:
  name: nginx-singleton-deployment
  labels:
    app.kubernetes.io/name: nginx-singleton
spec:
  replicas: 1
  selector:
    matchLabels:
      app: nginx
  template:
    metadata:
      labels:
        app: nginx
    spec:
      containers:
      - name: nginx
        image: public.ecr.aws/nginx/nginx:latest
        ports:
        - containerPort: 80
EOF

Verify that the status is available in wds1 for the deployment by running the command:

kubectl --context wds1 get deployments nginx-singleton-deployment -o yaml

Finally, scale the deployment from 1 to 2 replicas in wds1:

kubectl --context wds1 scale deployment nginx-singleton-deployment --replicas=2

and verify that replicas has been updated in cluster1 and wds1:

kubectl --context cluster1 get deployment nginx-singleton-deployment
kubectl --context wds1 get deployment nginx-singleton-deployment

[Optional] Teardown Scenario 4#

kubectl --context wds1 delete bindingpolicies nginx-singleton-bpolicy
kubectl --context wds1 delete deployments nginx-singleton-deployment

Scenario 5 - Resiliency testing#

This is a test that you can do after finishing Scenario 1.

TODO: rewrite this so that it makes sense after Scenario 4.

Bring down the control plane: stop and restart wds1 and its1 API servers, KubeFlex and KubeStellar controllers:

First stop all:

kubectl --context kind-kubeflex scale deployment -n wds1-system kube-apiserver --replicas=0
kubectl --context kind-kubeflex scale statefulset -n its1-system vcluster --replicas=0
kubectl --context kind-kubeflex scale deployment -n kubeflex-system kubeflex-controller-manager --replicas=0
kubectl --context kind-kubeflex scale deployment -n wds1-system kubestellar-controller-manager --replicas=0
kubectl --context kind-kubeflex scale deployment -n wds1-system transport-controller --replicas=0

Then restart all:

kubectl --context kind-kubeflex scale deployment -n wds1-system kube-apiserver --replicas=1
kubectl --context kind-kubeflex scale statefulset -n its1-system vcluster --replicas=1
kubectl --context kind-kubeflex scale deployment -n kubeflex-system kubeflex-controller-manager --replicas=1
kubectl --context kind-kubeflex scale deployment -n wds1-system kubestellar-controller-manager --replicas=1
kubectl --context kind-kubeflex scale deployment -n wds1-system transport-controller --replicas=1

Wait for about a minute for all pods to restart, then apply a new BindingPolicy:

kubectl --context wds1 apply -f - <<EOF
apiVersion: control.kubestellar.io/v1alpha1
kind: BindingPolicy
metadata:
  name: nginx-res-bpolicy
spec:
  clusterSelectors:
  - matchLabels: {"location-group":"edge"}
  downsync:
  - objectSelectors:
    - matchLabels: {"app.kubernetes.io/name":"nginx-res"}
EOF

and a new workload:

kubectl --context wds1 apply -f - <<EOF
apiVersion: v1
kind: Namespace
metadata:
  labels:
    app.kubernetes.io/name: nginx-res
  name: nginx-res
---
apiVersion: apps/v1
kind: Deployment
metadata:
  name: nginx-res-deployment
  namespace: nginx-res
  labels:
    app.kubernetes.io/name: nginx-res
spec:
  replicas: 1
  selector:
    matchLabels:
      app: nginx-res
  template:
    metadata:
      labels:
        app: nginx-res
    spec:
      containers:
      - name: nginx-res
        image: public.ecr.aws/nginx/nginx:latest
        ports:
        - containerPort: 80
EOF

Verify that deployment has been created in both clusters

kubectl --context cluster1 get deployments -n nginx-res
kubectl --context cluster2 get deployments -n nginx-res

[Optional] Teardown Scenario 5#

kubectl --context wds1 delete ns nginx-res
kubectl --context wds1 delete bindingpolicies nginx-res-bpolicy

Then continue with teardown of scenario 1.

Scenario 6 - multi-cluster workload deployment of app with ServiceAccount with ArgoCD#

This scenario is something you can do after the common setup.

Before running this scenario, install ArgoCD on the hosting cluster and configure it work with the WDS as outlined here.

Including a ServiceAccount tests whether there will be a controller fight over a token Secret for that ServiceAccount, which was observed in some situations with older code.

Apply the following BindingPolicy to wds1:

kubectl --context wds1 apply -f - <<EOF
apiVersion: control.kubestellar.io/v1alpha1
kind: BindingPolicy
metadata:
  name: argocd-sa-bpolicy
spec:
  clusterSelectors:
  - matchLabels: {"location-group":"edge"}
  downsync:
  - objectSelectors:
    - matchLabels: {"argocd.argoproj.io/instance":"nginx-sa"}
EOF

Switch context to hosting cluster and argocd namespace (this is required by argo to create an app with the CLI)

kubectl config use-context kind-kubeflex
kubectl config set-context --current --namespace=argocd

Create a new application in ArgoCD:

argocd app create nginx-sa --repo https://github.com/pdettori/sample-apps.git --path nginx --dest-server https://wds1.wds1-system --dest-namespace nginx-sa

Open browser to Argo UI:

open https://argocd.localtest.me:9443

Open the app nginx-sa and sync it by clicking the "sync" button and then "synchronize".

Alternatively, use the CLI to sync the app:

argocd app sync nginx-sa

Finally, check if the app has been deployed to the two clusters.

kubectl --context cluster1 -n nginx-sa get deployments,sa,secrets
kubectl --context cluster2 -n nginx-sa get deployments,sa,secrets

Repeat multiple syncing on Argo and verify that extra secrets for the service acccount are not created both wds1 and clusters:

kubectl --context wds1 -n nginx-sa get secrets
kubectl --context cluster1 -n nginx-sa get secrets
kubectl --context cluster2 -n nginx-sa get secrets

[Optional] Teardown Scenario 6#

(Assuming that kubectl is still using context kind-kubeflex and namespace argocd.)

argocd app delete nginx-sa --cascade
kubectl --context wds1 delete bindingpolicies argocd-sa-bpolicy