Operator SDK FAQ

What are the the differences between Kubebuilder and Operator-SDK?

Kubebuilder and Operator SDK are both projects that allow you to quickly create and manage an operator project. Operator SDK uses Kubebuilder under the hood to do so for Go projects, such that the operator-sdk CLI tool will work with a project created by kubebuilder. Therefore each project makes use of controller-runtime and will have the same basic layout. For further information also check the SDK Project Layout.

Operator SDK offers additional features on top of the basic project scaffolding that Kubebuilder provides. By default, operator-sdk init generates a project integrated with:

  • Operator Lifecycle Manager, an installation and runtime management system for operators
  • OperatorHub, a community hub for publishing operators
  • Operator SDK scorecard, a tool for ensuring operator best-practices and developing cluster tests

Operator SDK supports operator types other than Go as well, such as Ansible and Helm.

For further context about the relationship between Kubebuilder and Operator SDK, see this blog post.

Can I use the Kubebuilder docs?

Yes, you can use https://book.kubebuilder.io/. Just keep in mind that when you see an instruction such as: $ kubebuilder <command> you will use $ operator-sdk <command>.

Controller Runtime FAQ

Please see the upstream Controller Runtime FAQ first for any questions related to runtime mechanics or controller-runtime APIs.

Can I customize the projects initialized with operator-sdk?

After using the CLI to create your project, you are free to customize based on how you see fit. Please note that it is not recommended to deviate from the proposed layout unless you know what you are doing.

For example, you should refrain from moving the scaffolded files, doing so will make it difficult to upgrade your project in the future. You may also lose the ability to use some of the CLI features and helpers. For further information on the project layout, see the doc Project Layout

How can I have separate logic for Create, Update, and Delete events? When reconciling an object can I access its previous state?

You should not have separate logic. Instead design your reconciler to be idempotent. See the controller-runtime FAQ for more details.

How do I wait for some specific cluster state such as a resource being deleted?

You don’t. Instead, design your reconciler to be idempotent by taking the next step based on the current state, and then returning and requeuing. For example, waiting for an object to be deleted might look something like this:

func (r *Reconciler) Reconcile(ctx context.Context, req ctrl.Request) (res ctrl.Result, err error) {
    ...
    if !r.IfPodWasDeleted(ctx, pod) {
        if err := r.Delete(ctx, pod); err != nil {
            return ctrl.Result{}, err
        }
        return ctrl.Result{Requeue: true}, nil
    }
    // This code will be invoked only after pod deletion    
    r.DeployBiggerPod(ctx)
    ...
}

When my Custom Resource is deleted, I need to know its contents or perform cleanup tasks. How can I do that?

Use a finalizer.

I see the warning in my Operator’s logs: The resourceVersion for the provided watch is too old. What’s wrong?

This is completely normal and expected behavior.

The kube-apiserver watch request handler is designed to periodically close a watch to spread out load among controller node instances. Once disconnected, your Operator’s informer will automatically reconnect and re-establish the watch. If an event is missed during re-establishment, the watch will fail with the above warning message. The Operator’s informer then does a list request and uses the new resourceVersion from that list to restablish the watch and replace the cache with the latest objects.

This warning should not be stifled. It ensures that the informer is not stuck or wedged.

Never seeing this warning may suggest that your watch or cache is not healthy. If the message is repeating every few seconds, this may signal a network connection problem or issue with etcd.

For more information on kube-apiserver request timeout options, see the Kubernetes API Server Command Line Tool Reference

My Ansible module is missing a dependency. How do I add it to the image?

Unfortunately, adding the entire dependency tree for all Ansible modules would be excessive. Fortunately, you can add it easily. Simply edit your build/Dockerfile. You’ll want to change to root for the install command, just be sure to swap back using a series of commands like the following right after the FROM line.

USER 0
RUN yum -y install my-dependency
RUN pip3 install my-python-dependency
USER 1001

If you aren’t sure what dependencies are required, start up a container using the image in the FROM line as root. That will look something like this:

docker run -u 0 -it --rm --entrypoint /bin/bash quay.io/operator-framework/ansible-operator:<sdk-tag-version>

After deploying my operator, I see errors like “Failed to watch

If you run into the following error message, it means that your operator is unable to watch the resource:

E0320 15:42:17.676888       1 reflector.go:280] pkg/mod/k8s.io/client-go@v0.0.0-20191016111102-bec269661e48/tools/cache/reflector.go:96: Failed to watch *v1.ImageStreamTag: unknown (get imagestreamtags.image.openshift.io)
{"level":"info","ts":1584718937.766342,"logger":"controller_memcached","msg":"ImageStreamTag resource not found.

Using controller-runtime’s split client means that read operations (gets and lists) are read from a cache, and write operations are written directly to the API server. To populate the cache for reads, controller-runtime initiates a list and then a watch even when your operator is only attempting to get a single resource. The above scenario occurs when the operator does not have an RBAC permission to watch the resource. The solution is to add an RBAC directive to generate a config/rbac/role.yaml with watch privileges:

//+kubebuilder:rbac:groups=some.group.com,resources=myresources,verbs=watch

Alternatively, if the resource you’re attempting to cannot be watched (like v1.ImageStreamTag above), you can specify that objects of this type should not be cached by adding the following to main.go:

import (
	...
	imagev1 "github.com/openshift/api/image/v1"
)

var (
	scheme = runtime.NewScheme()
)

func init() {
	utilruntime.Must(clientgoscheme.AddToScheme(scheme))
	// Add imagev1's scheme.
	utilruntime.Must(imagev1.AddToScheme(scheme))
}

func main() {
	mgr, err := ctrl.NewManager(ctrl.GetConfigOrDie(), ctrl.Options{
		Scheme:        scheme,
		// Specify that ImageStreamTag's should not be cached.
		ClientBuilder: manager.NewClientBuilder().WithUncached(&imagev1.ImageStreamTag{}),
	})
}

Then in your controller file, add an RBAC directive to generate a config/rbac/role.yaml with get privileges:

//+kubebuilder:rbac:groups=image.openshift.io,resources=imagestreamtags,verbs=get

Now run make manifests to update your role.yaml.

After deploying my operator, why do I see errors like “is forbidden: cannot set blockOwnerDeletion if an ownerReference refers to a resource you can’t set finalizers on: …”?

If you are facing this issue, it means that the operator is missing the required RBAC permissions to update finalizers on the APIs it manages. This permission is necessary if the OwnerReferencesPermissionEnforcement plugin is enabled in your cluster.

For Helm and Ansible operators, this permission is configured by default. However for Go operators, it may be necessary to add this permission yourself by adding an RBAC directive to generate a config/rbac/role.yaml with update privileges on your CR’s finalizers:

//+kubebuilder:rbac:groups=cache.example.com,resources=memcacheds/finalizers,verbs=update

Now run make manifests to update your role.yaml.

When invoking make targets, why do I see errors like fork/exec /usr/local/kubebuilder/bin/etcd: no such file or directory occurred?

If using an OS or distro that does not point sh to the bash shell (Ubuntu for example), add the following line to the Makefile:

SHELL := /bin/bash

How do I make my Operator proxy-friendly?


Administrators can configure proxy-friendly Operators to support network proxies by specifying HTTP_PROXY, HTTPS_PROXY, and NO_PROXY environment variables in the Operator deployment. (These variables can be handled by OLM.)

Proxy-friendly Operators are responsible for inspecting the Operator environment and passing these variables along to the rquired operands. For more information and examples, please see the type-specific docs:

After running make manifests, rbac permissions are not updated in config

RBAC markers that are not followed by a newline will not be parsed correctly, resulting in missing rbac configuration.

This is a known issue with controller-tools, see issue #551 The current workaround is to add a new line after the rbac marker.

// +kubebuilder:rbac:groups=apps,resources=deployments,verbs=get;list;watch;create;update;patch;delete
// +kubebuilder:rbac:groups=core,resources=pods,verbs=get;list;

func (r *MemcachedReconciler) Reconcile(ctx context.Context, req ctrl.Request) (ctrl.Result, error) {

Preserve the preserveUnknownFields in your CRDs

The preserveUnknownFields will be removed if set to false when running make bundle. Because of some underlying data structure changes and how yaml is unmarshalled, it is best to add them back in after they have been written.

You can use this script to post process the files to add the preserveUnknownFields back in.

function generate_preserveUnknownFieldsdata() {
    for j in config/crd/patches/*.yaml ; do
        if grep -qF "preserveUnknownFields" "$j";then
            variable=`awk '/metadata/{flag=1} flag && /name:/{print $NF;flag=""}' "$j"`
            for k in config/crd/bases/*.yaml ; do
                if grep -qF "$variable" "$j";then
                    filename=`awk 'END{ var=FILENAME; split (var,a,/\//); print a[4]}' "$k"`
                    awk '/^spec:/{print;print "  preserveUnknownFields: false";next}1' "bundle/manifests/$filename" > testfile.tmp && mv testfile.tmp "bundle/manifests/$filename"
                fi
            done
        fi
    done
}

generate_preserveUnknownFieldsdata

You can then modify the bundle target in your Makefile by adding a call to the script at the end of the target. See the example below:

.PHONY: bundle
bundle: manifests kustomize ## Generate bundle manifests and metadata, then validate generated files.
	operator-sdk generate kustomize manifests -q
	cd config/manager && $(KUSTOMIZE) edit set image controller=$(IMG)
	$(KUSTOMIZE) build config/manifests | operator-sdk generate bundle -q --overwrite --version $(VERSION) $(BUNDLE_METADATA_OPTS)
	operator-sdk bundle validate ./bundle | ./preserve_script.sh

Note: Though this is a bug with controller-gen which is used by Operator SDK to generate CRD, this is a workaround from our end to enable users to preserve the field after controller-gen has run.

What is the bundle limit size? Was this amount increased?

Bundles have a size limitation because their manifests are used to create a configMap, and the Kubernetes API does not allow configMaps larger than ~1MB. Beginning with OLM version v0.19.0 and OPM 1.17.5, these values are now compressed accommodating larger bundles. (More info).

The change to allow bigger bundles from OLM version v0.19.0 only impacts the full bundle size amount. Any single manifest within the bundle such as the CRD will still make the bundle uninstallable if it exceeds the default file size limit on clusters (~1MB).

The size of my Operator bundle is too big. What can I do?

If your bundle is too large, there are a few things you can try:

  • Reducing the number of CRD versions supported in your Operator by deprecating and then removing older API versions. It is a good idea to have a clear plan for deprecation and removal of old CRDs versions when new ones get added, see Kubernetes API change practices. Also, refer to the Kubernetes API conventions.
  • Reduce the verbosity of your API documentation. (We do not recommend eliminating documenting the APIs)

How can I update dependencies for an unsupported release image?

The Operator-SDK community releases updated images for supported releases. If you are using an older version of Operator-SDK, sometimes the dependencies will need to be updated in the images. For users in this situation we recommend updating to the latest version. If this is not possible, users can build and push their own versions of any of the images provided by the Operator-SDK.

Operator-SDK docker buildx build -t quay.io/operator-framework/operator-sdk:dev -f ./images/operator-sdk/Dockerfile –load .

Helm-Operator docker buildx build -t quay.io/operator-framework/helm-operator:dev -f ./images/helm-operator/Dockerfile –load .

Scorecard-test docker buildx build -t quay.io/operator-framework/scorecard-test:dev -f ./images/scorecard-test/Dockerfile –load .

Scorecard-test-kuttl docker buildx build -t quay.io/operator-framework/scorecard-test-kuttl:dev -f ./images/scorecard-test-kuttl/Dockerfile –load .

Ansible

Ansible images are built in 2 layers, and both will need to be rebuilt. Build and push the dependency image images/ansible-operator/base.Dockerfile, and then update FROM in images/ansible-operator/Dockerfile to point to your image, and build and push this image, which can be added to your operator’s FROM.

Ansible Operator (2.9) base docker buildx build -t quay.io/operator-framework/ansible-operator-base:dev -f ./images/ansible-operator/base.Dockerfile --load images/ansible-operator

Ansible Operator (2.9) docker buildx build -t quay.io/operator-framework/ansible-operator:dev -f ./images/ansible-operator/Dockerfile --load .

Ansible Operator (2.11) Dependencies docker buildx build -t quay.io/operator-framework/ansible-operator-2.11-preview-base:dev -f ./images/ansible-operator-2.11-preview/base.Dockerfile --load images/ansible-operator-2.11-preview

Ansible Operator (2.11) docker buildx build -t quay.io/operator-framework/ansible-operator-2.11-preview:dev -f ./images/ansible-operator-2.11-preview/Dockerfile --load .

Running operator-sdk create api results in an error with /usr/local/go/src/net/cgo_linux.go:13:8: no such package located in the error message

By default Go will set the CGO_ENABLED environment variable to 1 which means that cgo is enabled. Depending on the architecture and OS of your system you may run into an issue similar to this one:

/usr/local/go/src/net/cgo_linux.go:13:8: no such package located
Error: not all generators ran successfully
run `controller-gen object:headerFile=hack/boilerplate.go.txt paths=./... -w` to see all available markers, or `controller-gen object:headerFile=hack/boilerplate.go.txt paths=./... -h` for usage
make: *** [Makefile:95: generate] Error 1
Error: failed to create API: unable to run post-scaffold tasks of "base.go.kubebuilder.io/v3": exit status 2

Here are a couple workarounds to try to resolve the issue:

  • Ensure gcc is installed
  • Set the CGO_ENABLED environment variable to 0 to disable cgo

If neither of those solutions work for you, please open an issue


Last modified April 13, 2022: (docs): update Go prereqs (#5650) (75eb77e8)