Install Openshift Container Storage (OCS) on OpenShift infrastructure nodes

Using Infrastructure nodes to schedule OpenShift Container Storage (OCS) resources will save on OpenShift Container Platform (OCP) subscription costs. Any OCP node that has a infra node-role label will only require OCS subscription but no OCP subscription. # Background Currently the Machine API cannot handle the creation of nodes carrying only the node-role.kubernetes.io/infra node-role label. Nodes created with the Machine API can only have node-roles added along with the default node-role.kubernetes.io/worker.

A common approach is desirable for consistency across environments, both those with and without Machine API support (reference section below for manual configuration of infa nodes). Because of this, it is highly recommended in all cases to have nodes with the dual worker/infra node-role or label. # Anatomy of an Infrastructure node Infrastructure nodes for use with OCS have a few attributes. The infra node-role label is required to ensure the node does not consume OCP entitlements. Effectively, the infra node-role label is responsible for ensuring only OCS entitlements are necessary for the nodes running OCS.

Adding a OCS taint with a NoSchedule effect is also required so that the infra node will only schedule OCS resources.

The label identifies the OCP node as a infra node so that OCP subscription cost is not applied. The taint prevents non-OCS resources to be scheduled on the tainted nodes. If using local storage devices for OCS then a toleration will need to be added to allow Local Storage Operator (LSO) resources to schedule on the infra nodes. Reference section below for how this is done.

Example of the taint and labels required on infrastructure node that will be used to run OCS services:

If the Machine API is supported in the environment, then labels should be added to the templates for the Machine Sets that will be provisioning the infrastructure nodes. Avoid the anti-pattern of adding labels manually to nodes created by the machine API. Doing so is analogous to adding labels to pods created by a deployment. In both cases, when the pod/node fails, the replacement pod/node will not have the appropriate labels.

Example Machine Set template that creates nodes with the appropriate taint and labels required for infrastructure nodes that will be used to run OCS services:

Only when the Machine APi is not supported in the environment should labels be directly applied to nodes. Manual creation will require that at least 3 OCP worker nodes are available to schedule OCS services and that these nodes have sufficient CPU and memory resources. To avoid the OCP subscription cost the following is required:

Adding a NoSchedule OCS taint is also required so that the infra node will only schedule OCS resources and repel any other non-OCS workloads.

The removal of the worker node-role is not recommended. If already removed, it should be added again to each infra node. Adding a infra node-role and OCS taint is sufficient to conform to entitlement exemption requirements, and it is not necessary to remove the worker node-role. In fact, removing the worker node-role from infra nodes can cause issues unless changes are made both to the OpenShift scheduler and to MachineConfig resources.

When local storage devices are used for creating the OCS cluster (i.e., AWS i3en.3xlarge instance type) then LSO will need to installed before OCS can be deployed. In order to allow the LSO pods to schedule on the infra nodes with the OCS NoSchedule taint, a toleration has to be added to the LocalVolume custom resource file.

Here is an example with the toleration node.ocs.openshift.io/storage=true added.

This will allow LSO pods to schedule and PVs to be created from devices listed under devicePaths:. The recommended practice is to use /dev/disk/by-id/ to identify the storage devices that will be used for OCS.