In the last few weeks, we had an opportunity to interact with number of Kubernetes users and community members as a result of some of the presentations we did at meetup events in San Antonio, Austin and Dallas. Here are the recurring questions that we heard on topics related to CRDs and Operators during these sessions.
Q. What is a Kubernetes Operator?
A. Think of an Operator as essentially a new REST API that is added to a Kubernetes cluster’s control plane. Like traditional REST APIs, it has a resource path, a resource definition and code that knows how to perform CRUD operations on that resource. In Kubernetes nomenclature, the resource path is called as the Custom Resource, resource definition is called as the Custom Resource Kind, and the code that performs CRUD operations is called Custom Controller. Most of the times no distinction is made between Custom Resource and Custom Resource Kind when discussing CRDs and Operators. ‘Custom Resource’ is used to refer to both. The Custom Controller typically implements domain-specific knowledge such as managing stateful services like databases on Kubernetes.
Q. Why are Operators useful?
A. Operators offer Kubernetes-native way to do your platform automation by extending Kubernetes for your platform workflows. It allows Operator developers to easily share their automation with the broader community enabling composability and reuse. Ultimately this approach makes it possible to reduce in-house custom platform automation and at the same time offers a declarative way to define the platform stacks using Kubernetes YAMLs. Check this post for details.
Q. How are Kubernetes Operators different than Helm charts?
A. Operators and Helm charts serve different purpose. Helm is a tool that simplifies deploying Kubernetes manifests to a cluster. It focuses on templatization, reusability, etc. of Kubernetes YAMLs. Helm does not handle CRUD operations on Custom Resources. On the other hand, Operators have knowledge about what it means to perform CRUD operations on Custom Resources. In fact, Helm is a very good mechanism to package and deploy Operators. Helm can also be used to deploy Kubernetes manifests that include Custom Resources introduced by Operators.
Q. What is the difference between an Operator and a Custom Controller?
A. Custom Controller which is not part of an Operator is something that understands only Kubernetes’s native abstractions such as Pod, Deployment, Service, etc. A Custom Controller that is part of an Operator also understands Custom Resource abstractions that the Operator has introduced (e.g. MysqlCluster).
Q. Where does an Operator run?
A. An Operator runs inside of a Kubernetes cluster.
Q. What is needed in Kubernetes to run an Operator?
A. To run an Operator, you need the ‘Custom Resource Definition (CRD)’ meta REST API enabled in your cluster. A CRD enables registering custom REST APIs into a Kubernetes cluster. If your Kubernetes cluster version is 1.9 or above then you already have this meta REST API enabled.
Q. How to run an Operator in a cluster?
A. To run an Operator, you need to first create a container image of your Operator code (essentially image that packages Custom Controller code). Then use a CRD object to register into your cluster’s set of APIs, the Custom Resource API that your Operator is managing. Then create a Deployment manifest with the container image that you built. Finally, use kubectl or helm to apply/install the Deployment manifest in the cluster.
Q. What permissions are required to deploy an Operator?
A. The CRD object is not namespaced. It is available at the Cluster scope. So in order to register CRD for your Operator, the user needs ClusterRole permission on the customresourcedefinition object. If in your setup regular users do not have Cluster scope permissions then the cluster admin will have to install the CRD object. Deploying the Operator’s container Pod requires Role level permission on the deployment object. This permission can be granted to a regular user. So the Operator Pod can be created by a regular user. It is easier if both steps — installing the CRD and installing Operator Pod — are done by the same user to simplify Operator deployment.
Q. What permissions are required for Operator Pods?
A. This depends on the type of actions that the Operator’s custom controller is designed to perform. If an Operator’s controller is creating Kubernetes resources that are Cluster-level then the Operator’s Pod will need corresponding ClusterRole permissions. For instance, if the controller is creating PersistentVolumes then the Pod will need Cluster-level permission for managing PersistentVolumes. You will need to create a Service Account and grant it such permissions. Then use this Service Account in your Operator Pod Spec.
Q. Can an Operator be deployed in a particular namespace?
A. Operator Pod can be deployed in any namespace. The CRD object is Cluster wide though.
Q. Can one Operator inherit another Operator?
A. No, there is no inheritance in the style of Object-Oriented languages between Operators.
Q. Are there any standards emerging around Operators and Custom Resources?
A. Not yet. The only standard right now is that the Custom Resource Spec definition should follow Kubernetes Spec definition format. Operator developers can define their own Custom Resource Spec properties.
Q. What is the purpose of Operators that provision Cloud Services like AWS RDS?
A. Such Operators can be used to provision Cloud-based managed services directly using kubectl.
Q. Will running MySQL on Kubernetes using a MySQL Operator provide same level of robustness as a managed database service like AWS RDS or Google CloudSQL?
A. It depends on how the Operator’s code is written. You can choose from multiple available MySQL community Operators or develop your own or customize a community Operator further for your internal requirements.
Q. Are there situations where one needs multiple Operators?
A. Often. We are seeing enterprises use multiple Operators to build their platform stacks on Kubernetes. At CloudARK, we have pioneered Platform-as-Code approach for creating multi Operator platform stacks.
Q. What are the advantages of using multiple Operators?
A. Development and maintenance of in-house platform automation is reduced. This approach helps you in your goal of moving towards multi-cloud application portability as this significantly simplifies creating or transferring your workloads from one cloud environment to other.
Q. What are the challenges when using multiple Operators together?
A. Primary challenge is interoperability and binding between various Custom Resources supported by Operators. We have developed KubePlus Platform Toolkit that helps with this.
Q. How to use KubePlus Platform Toolkit?
KubePlus Platform Toolkit standardizes on two things — annotations that can be added to Custom Resource Definitions of your Operators, and custom endpoints that can be used for Custom Resource discovery. Operator developer/curator defines the annotations with appropriate values. KubePlus Platform Toolkit uses these annotations to surface information about Custom Resources through the defined custom endpoints. This enables users of Custom Resources to understand static as well as dynamic information about them which can be used in figuring out their interoperability characteristics. In short, KubePlus Platform Kit offers additional endpoints to learn more information about Custom Resources in your cluster, specially that information which is hard to extract using standard Kubernetes interfaces like “describe” or “explain”.
Here are examples of new endpoints enabled by KubePlus Platform Toolkit:
Use ‘man’ endpoint to find usage information about Custom Resource:
e.g. kubectl get — raw “/apis/platform-as-code/v1/man?kind=Moodle”
Use ‘composition’ endpoint to find underlying resources that the custom resource is composed of:
e.g. kubectl get — raw “/apis/platform-as-code/v1/composition?kind=Moodle&instance=moodle1&namespace=ns1”
For a detailed example check this.
Q. What are all the personas involved in developing, installing, using Operators? What tools exist for each?
A. The three personas involved are — Operator developers/Operator curators, Cluster Admins/DevOps Engineers, Application developers. The tools available for each persona are listed in following table:
Q. In what language are Operators written? Is there a preferred language?
Q. We are interested in building our Kubernetes platform using Open source Operators. What should we look for when choosing an Operator?
A. There are Operator listing and repositories such as following: https://operatorhub.io/, https://chartmuseum.com/, https://github.com/operator-framework/awesome-operators, https://kubedex.com/operators/. You can pick an Operator from any of these places. Then use the guidelines that we have developed to evaluate whether the Operator that you have selected has some of the attributes mentioned in the guidelines.
At CloudARK, we help enterprises build platforms for their specific needs on top of Kubernetes leveraging Kubernetes CRDs and Operators. So feel free to contact us if you are interested in getting help on curating, customizing or developing Operators for your platform specific needs.
Q. Are Operators production ready? Is there any analysis of open source Operators?
A. We have done analysis of some of the open source Operators. You can find it here. The space of Operators is evolving very fast and most of the Operators are currently under active development. If you are interested in any particular Operator, let us know. We will evaluate it for your needs.
Q. Are there situations where Operator pattern cannot be used? Or, what is an Operator anti-pattern?
A. Operator pattern is not suitable if there is no need for monitoring and reconciling the cluster state based on some declarative input leveraging Custom Resource abstraction. Operator pattern is also not suitable if actions that need to be performed on the cluster can not be translated into declarative input and instead are imperative in nature by design. For handling such requirement, you can either create a separate Pod and run it in the cluster, or if you need the functionality through kubectl then consider using Kubernetes Aggregated API Server.
We plan to keep this FAQ as a working document. So if you have any questions about Kubernetes CRDs/Operators, feel free to add them in comments, or send them to us. We will include answers to your questions in the next iteration of the FAQ.
Also if you are interested to learn more about multi-Operator platform stacks, sign-up here to get your free copy of our Platform-as-Code eBook.