The operator consists of the following parts:
The operator is packaged in a container image which you can access using the following
docker pull commands:
$ docker pull ghcr.io/oracle/weblogic-kubernetes-operator:3.1.4
For more details on acquiring the operator image and prerequisites for installing the operator, consult the Quick Start guide.
The operator registers a Kubernetes custom resource definition called
domains). More details about the Domain type defined by this CRD, including its schema, are available here.
The diagram below shows the general layout of high-level components, including optional components, in a Kubernetes cluster that is hosting WebLogic domains and the operator:
The Kubernetes cluster has several namespaces. Components may be deployed into namespaces as follows:
The diagram below shows how the various parts of a WebLogic domain are manifest in Kubernetes by the operator.
This diagram shows the following details:
applicationsdirectory, a directory for storing logs, and a directory for any file-based persistence stores.
weblogic.domainName. One container runs in this pod. WebLogic Node Manager and Administration Server processes are run inside this container. The Node Manager process is used as an internal implementation detail for the liveness probe, for patching, and to provide monitoring and control capabilities to the Administration Console. It is not intended to be used for other purposes, and it may be removed in some future release.
ClusterIPtype service is created for the Administration Server pod. This service provides a stable, well-known network (DNS) name for the Administration Server. This name is derived from the
domainUIDand the Administration Server name, and it is known before starting up any pod. The Administration Server
ListenAddressis set to this well-known name.
ClusterIPtype services are only visible inside the Kubernetes cluster. They are used to provide the well-known names that all of the servers in a domain use to communicate with each other. This service is labeled with
NodePorttype service is optionally created for the Administration Server pod. This service provides HTTP access to the Administration Server to clients that are outside the Kubernetes cluster. This service is intended to be used to access the WebLogic Server Administration Console or for the T3 protocol for WLST connections. This service is labeled with
weblogic.domainName. One container runs in each pod. WebLogic Node Manager and Managed Server processes are run inside each of these containers. The Node Manager process is used as an internal implementation detail for the liveness probe. It is not intended to be used for other purposes, and it may be removed in some future release.
ClusterIPtype service is created for each Managed Server pod that contains a Managed Server that is not part of a WebLogic cluster. These services are intended to be used to access applications running on the Managed Servers. These services are labeled with
weblogic.domainName. Customers must expose these services using a load balancer or
NodePorttype service to expose these endpoints outside the Kubernetes cluster.
Kubernetes requires that the names of some resource types follow the DNS label standard as defined in DNS Label Names and RFC 1123. Therefore, the operator enforces that the names of the Kubernetes resources do not exceed Kubernetes limits (see Meet Kubernetes resource name restrictions).
The diagram below shows the components inside the containers running WebLogic Server instances:
The Domain specifies a container image, defaulting to
container-registry.oracle.com/middleware/weblogic:220.127.116.11. All containers running WebLogic Server use this same image. Depending on the use case, this image could contain the WebLogic Server product binaries or also include the domain directory.
During a rolling event caused by a change to the Domain’s
image field, containers will be using a mix of the updated value of the
image field and its previous value.
ENTRYPOINTis configured by a script that starts up a Node Manager process, and then uses WLST to request that Node Manager start the server. Node Manager is used to start servers so that the socket connection to the server will be available to obtain server status even when the server is unresponsive. This is used by the liveness probe.
The operator expects (and requires) that all state be stored outside of the images that are used to run the domain. This means either in a persistent file system, or in a database. The WebLogic configuration, that is, the domain directory and the applications directory may come from the image or a persistent volume. However, other state, such as file-based persistent stores, and such, must be stored on a persistent volume or in a database. All of the containers that are participating in the WebLogic domain use the same image, and take on their personality; that is, which server they execute, at startup time. Each Pod mounts storage, according to the Domain, and has access to the state information that it needs to fulfill its role in the domain.
It is worth providing some background information on why this approach was adopted, in addition to the fact that this separation is consistent with other existing operators (for other products) and the Kubernetes “cattle, not pets” philosophy when it comes to containers.
The external state approach allows the operator to treat the images as essentially immutable, read-only, binary images. This means that the image needs to be pulled only once, and that many domains can share the same image. This helps to minimize the amount of bandwidth and storage needed for WebLogic Server images.
This approach also eliminates the need to manage any state created in a running container, because all of the state that needs to be preserved is written into either the persistent volume or a database backend. The containers and pods are completely throwaway and can be replaced with new containers and pods, as necessary. This makes handling failures and rolling restarts much simpler because there is no need to preserve any state inside a running container.
When users wish to apply a binary patch to WebLogic Server, it is necessary to create only a single new, patched image. If desired, any domains that are running may be updated to this new patched image with a rolling restart, because there is no state in the containers.
It is envisaged that in some future release of the operator, it will be desirable to be able to “move” or “copy” domains in order to support scenarios like Kubernetes federation, high availability, and disaster recovery. Separating the state from the running containers is seen as a way to greatly simplify this feature, and to minimize the amount of data that would need to be moved over the network, because the configuration is generally much smaller than the size of WebLogic Server images.
The team developing the operator felt that these considerations provided adequate justification for adopting the external state approach.
The operator uses services to provide stable, well-known DNS names for each server. These names are known in advance of starting up a pod to run a server, and are used in the
ListenAddress fields in the WebLogic Server configuration to ensure that servers will always be able to find each other. This also eliminates the need for pod names or the actual WebLogic Server instance names to be the same as the DNS addresses.