Migrating Legacy Projects

    The motivations for the new layout are related to bringing more flexibility to users and part of the process to integrate Kubebuilder and Operator SDK.

    • The deploy directory was replaced with the config directory including a new layout of Kubernetes manifests files:

      • CRD manifests in deploy/crds/ are now in config/crd/bases
      • CR manifests in deploy/crds/ are now in config/samples
      • Controller manifest deploy/operator.yaml is now in config/manager/manager.yaml
      • RBAC manifests in deploy are now in config/rbac/

    • build/Dockerfile is moved to Dockerfile in the project root directory

    • pkg/apis and pkg/controllers are now in the root directory.

    • cmd/manager/main.go is now in the root directory.

    What is new

    Projects are now scaffold using:

    • kustomize to manage Kubernetes resources needed to deploy your operator
    • A Makefile with helpful targets for build, test, and deployment, and to give you flexibility to tailor things to your project’s needs
    • Helpers and options to work with webhooks
    • Updated metrics configuration using , a --metrics-addr flag, and kustomize-based deployment of a Kubernetes Service and prometheus operator ServiceMonitor
    • Scaffolded tests that use the test framework

    Migration Steps

    The most straightforward migration path is to:

    1. Create a new project from scratch to let operator-sdk scaffold the new project.
    2. Copy your existing code and configuration into the new project structure.

    Note: It is recommended that you have your project upgraded to the latest SDK release version before following the steps of this guide to migrate to new layout.

    In Kubebuilder-style projects, CRD groups are defined using two different flags (--group and --domain).

    When we initialize a new project, we need to specify the domain that all APIs in our project will share, so before creating the new project, we need to determine which domain we’re using for the APIs in our existing project.

    To determine the domain, look at the spec.group field in your CRDs in the deploy/crds directory.

    The domain is everything after the first DNS segment. Using cache.example.com as an example, the --domain would be example.com.

    So let’s create a new project with the same domain (example.com):

    Note: operator-sdk attempts to automatically discover the Go module path of your project by looking for a go.mod file, or if in $GOPATH, by using the directory path. Use the --repo flag to explicitly set the module path.

    Check if your project is multi-group

    Before we start to create the APIs, check if your project has more than one group such as : foo.example.com/v1 and crew.example.com/v1. If you intend to still working on with multiple groups in your project, then add the line multigroup: true in the PROJECT file. The PROJECT file for the above example would look like:

    1. domain: example.com
    2. repo: github.com/example-inc/memcached-operator
    3. multigroup: true
    4. version: 2
    5. ...

    Note: In multi-group projects, APIs are defined in apis/<group>/<version> and controllers are defined in controllers/<group>.

    Migrate APIs and Controllers

    Now that we have our new project initialized, we need to re-create each of our APIs. Using our API example from earlier (cache.example.com), we’ll use cache for the --group, v1alpha1 for the --version and Memcached for --kind flag.

    1. operator-sdk create api \
    2.     --group=cache \
    3.     --version=<version> \
    4.     --kind=<Kind> \
    5.     --controller

    API’s

    Now let’s copy the API definition from pkg/apis/<group>/<version>/<kind>_types.go to api/<version>/<kind>_types.go. For our example, it is only required to copy the code from the Spec and Status fields.

    This file is quite similar to the old one. Once you copy over your API definitions and generate manifests, you should end up with an identical API for your custom resource type. However, pay close attention to these kubebuilder Markers:

    • The +k8s:deepcopy-gen:interfaces=... marker was replaced with +kubebuilder:object:root=true.
    • If you are not using to generate OpenAPI Go code, then // +k8s:openapi-gen=true and other related openapi markers can be removed.

    NOTE The operator-sdk generate openapi command was deprecated in 0.13.0 and was removed from SDK release version. So far, it is recommended to use openapi-gen directly for OpenAPI code generation.

    Our Memcached API types will look like:

    1. // MemcachedSpec defines the desired state of Memcached
    2. type MemcachedSpec struct {
    3.     // Size is the size of the memcached deployment
    4.     Size int32 `json:"size"`
    5. }
    7. // MemcachedStatus defines the observed state of Memcached
    8. type MemcachedStatus struct {
    9.     // Nodes are the names of the memcached pods
    10.     Nodes []string `json:"nodes"`
    11. }
    13. // +kubebuilder:object:root=true
    14. // +kubebuilder:subresource:status
    16. // Memcached is the Schema for the memcacheds API
    17. type Memcached struct {...}
    19. // +kubebuilder:object:root=true
    21. // MemcachedList contains a list of Memcached
    22. type MemcachedList struct {...}

    Now let’s migrate the controller code from pkg/controller/<kind>/<kind>_controller.go to controllers/<kind>_controller.go. Following the steps:

    1. Copy over any struct fields from the existing project into the new <Kind>Reconciler struct. Note The Reconciler struct has been renamed from Reconcile<Kind> to <Kind>Reconciler. In our example, we would see ReconcileMemcached instead of MemcachedReconciler.
    2. Replace the // your logic here in the new layout with your reconcile logic.
    3. Copy the code under func add(mgr manager.Manager, r reconcile.Reconciler) to func SetupWithManager:
    1. func (r *MemcachedReconciler) SetupWithManager(mgr ctrl.Manager) error {
    2.     return ctrl.NewControllerManagedBy(mgr).
    3.         For(&cachev1alpha1.Memcached{}).
    4.         Owns(&appsv1.Deployment{}).
    5.         Complete(r)
    6. }

    In our example, the Watch implemented for the Deployment will be replaced with Owns(&appsv1.Deployment{}). Setting up controller Watches is simplified in more recent versions of controller-runtime, which has controller helpers to handle more of the details.

    Set the RBAC permissions

    The RBAC permissions are now configured via , which are used to generate and update the manifest files present in config/rbac/. These markers can be found (and should be defined) on the Reconcile() method of each controller.

    In the Memcached example, they look like the following:

    To update config/rbac/role.yaml after changing the markers, run make manifests.

    By default, new projects are cluster-scoped (i.e. they have cluster-scoped permissions and watch all namespaces). Read the operator scope documentation for more information about changing the scope of your operator.

    See the complete migrated memcached_controller.go code .

    By checking our new main.go we will find that:

    • The SDK leader.Become was replaced by the with lease mechanism. However, you still able to stick with the leader.Become for life if you wish:
    1. func main() {
    2. ...
    3.     ctx := context.TODO()
    4.     // Become the leader before proceeding
    5.     err = leader.Become(ctx, "memcached-operator-lock")
    6.     if err != nil {
    7.         log.Error(err, "")
    8.         os.Exit(1)
    9.     }

    In order to use the previous one ensure that you have the as a dependency of your project.

    • The default port used by the metric endpoint binds to :8080 from the previous :8383. To continue using port 8383, specify --metrics-addr=:8383 when you start the operator.

    • OPERATOR_NAME and POD_NAME environment variables are no longer used. OPERATOR_NAME was used to define the name for a leader election config map. Operator authors should use the LeaderElectionID attribute from the Manager Options which is hardcoded in :

    1. func main() {
    2. ...
    3.     mgr, err := ctrl.NewManager(ctrl.GetConfigOrDie(), ctrl.Options{
    4.         Scheme:             scheme,
    5.         MetricsBindAddress: metricsAddr,
    6.         Port:               9443,
    7.         LeaderElection:     enableLeaderElection,
    8.         LeaderElectionID:   "f1c5ece8.example.com",
    9.     })
    10. ...
    • Ensure that you copy all customizations made in cmd/manager/main.go to main.go. You’ll also need to ensure that all needed schemes have been registered, if you have been using third-party API’s (i.e Route Api from OpenShift).

    Migrate your tests

    For the new layout, you will see that controllers/suite_test.go is created when a controller is scaffolded by the tool. This file contains boilerplate for executing integration tests using envtest with and gomega.

    Operator SDK 1.0.0+ removes support for the legacy test framework and no longer supports the operator-sdk test subcommand. All affected tests should be migrated to use envtest.

    To learn more about how you can test your controllers, see the documentation about .

    Migrate your Custom Resources

    Custom resource samples are stored in ./config/samples in the new project structure. Copy the examples from your existing project into this directory. In existing projects, CR files have the format ./deploy/crds/<group>.<domain>_<version>_<kind>_cr.yaml.

    In our example, we’ll copy the specs from deploy/crds/cache.example.com_v1alpha1_memcached_cr.yaml to config/samples/cache_v1alpha1_memcached.yaml

    Configure your Operator

    In case your project has customizations in the deploy/operator.yaml then, it needs to be port to config/manager/manager.yaml. Note that, OPERATOR_NAME and POD_NAME env vars are no longer used. For further information came back to the section Migrate main.go.

    If you are using metrics and would like to keep them exported, see that the func addMetrics() is no longer generated in the main.go and it is now configurable via . Following the steps.

    • Ensure that you have Prometheus installed in the cluster: To check if you have the required API resource to create the ServiceMonitor run:
    1. kubectl api-resources | grep servicemonitors

    If not, you can install Prometheus via kube-prometheus:

    1. kubectl apply -f https://raw.githubusercontent.com/coreos/prometheus-operator/release-0.33/bundle.yaml
    • Now uncomment the line - ../prometheus in the config/default/kustomization.yaml file. It creates the ServiceMonitor resource which enables exporting the metrics:

    Use Handler from operator-lib

    By using the InstrumentedEnqueueRequestForObject you will able to export metrics from your Custom Resources. In our example, it would like:

    1. import (
    2.     ...
    3.     "github.com/operator-framework/operator-lib/handler"
    4.     ...
    5. )
    7. func (r *MemcachedReconciler) SetupWithManager(mgr ctrl.Manager) error {
    8.     // Create a new controller
    9.     c, err := controller.New("memcached-controller", mgr, controller.Options{Reconciler: r})
    10.     if err != nil {
    11.         return err
    12.     }
    13.     ...
    14.     err = c.Watch(&source.Kind{Type: &cachev1alpha1.Memcached{}}, &handler.InstrumentedEnqueueRequestForObject{})
    15.     if err != nil {
    16.         return err
    17.     }
    18.     ...
    19.     return nil
    20. }

    Note Ensure that you have the added to your go.mod.

    In this way, the following metric with the resource info will be exported:

    1. resource_created_at_seconds{"name", "namespace", "group", "version", "kind"}

    Note: To check it you can create a pod to curl the metrics/ endpoint but note that it is now protected by the kube-auth-proxy which means that you will need to create a ClusterRoleBinding and obtained the token from the ServiceAccount’s secret which will be used in the requests. Otherwise, to test you can disable the as well.

    For more info see the metrics.

    Operator image

    The Dockerfile image also changes and now it is a multi-stage, distroless and still been rootless, however, users can change it to work as however they want.

    See that, you might need to port some customizations made in your old Dockerfile as well. Also, if you wish to still using the previous UBI image replace:

    1. # Use distroless as minimal base image to package the manager binary
    2. # Refer to https://github.com/GoogleContainerTools/distroless for more details
    3. FROM gcr.io/distroless/static:nonroot

    With:

    1. FROM registry.access.redhat.com/ubi8/ubi-minimal:latest

    Generate Manifests and Build the operator

    Note that:

    • operator-sdk generate crds is replaced with make manifests, which generates CRDs and RBAC rules.
    • operator-sdk build is replaced with make docker-build IMG=<some-registry>/<project-name>:tag.

    In this way, run:

    Verify the migration

    The project can now be built, and the operator can be deployed on-cluster. For further steps regarding the deployment of the operator, creation of custom resources, and cleaning up of resources, see the quickstart guide.