CNCF Atlas

English

日本語

English

日本語

Appearance

Getting Started ​

Verified against tag v0.2.0 (commit 761a00b). Commands assume a working kubectl context, Go 1.24+, Docker, and a container registry you can push to.

Prerequisites ​

  • Go version v1.24.0+ (the README and go.mod pin go 1.24.0).
  • Docker version 17.03+.
  • kubectl version v1.11.3+ and access to a Kubernetes v1.11.3+ cluster.
  • For a real attach: a supported CDI provider (Fujitsu FTI_CDI, SNIA Sunfish, or NEC) and the composable fabric behind it. Without one, you can install and run the operator, but ComposabilityRequest objects will not complete.

Install ​

Clone the repository and check out the documented version.

bash
git clone https://github.com/CoHDI/composable-resource-operator.git
cd composable-resource-operator
git checkout v0.2.0

Build the manager binary to confirm the toolchain is set up. The build target runs manifests, generate, fmt, and vet first.

bash
make build

A first working setup ​

  1. Install the CRDs into the cluster your kubeconfig points at.

    bash
    make install

  2. Confirm both CRDs registered.

    bash
    kubectl get crds | grep cro.hpsys.ibm.ie.com

  3. Build and push the operator image to your registry.

    bash
    make docker-build docker-push IMG=<some-registry>/composable-resource-operator:v0.2.0

  4. Deploy the manager. The deployment needs a provider configured through environment variables (CDI_PROVIDER_TYPE, DEVICE_RESOURCE_TYPE, and provider endpoints); edit config/manager before applying if you are wiring real hardware.

    bash
    make deploy IMG=<some-registry>/composable-resource-operator:v0.2.0

  5. Create a ComposabilityRequest. This is the spec form from the project README.

    bash
    kubectl apply -f - <<'EOF'
apiVersion: cro.hpsys.ibm.ie.com/v1alpha1
kind: ComposabilityRequest
metadata:
  name: composabilityrequest-sample
spec:
  resource:
    type: "gpu"
    size: 2
    model: "NVIDIA-A100-PCIE-40GB"
    target_node: "node1"
    allocation_policy: "samenode"
EOF

Verify it works ​

Check that the operator is running and that the request is progressing through its state machine.

bash
kubectl get pods -n composable-resource-operator-system
kubectl get composabilityrequest composabilityrequest-sample -o jsonpath='{.status.state}'

With a configured fabric the state moves NodeAllocating to Updating to Running, and one ComposableResource per device appears.

bash
kubectl get composableresource

Without a fabric, the request stays in an earlier state; that still confirms the CRDs and controller are installed and reconciling.

Where to go next ​

  • Uninstall with make undeploy and make uninstall when you are done.
  • The Architecture and Internals pages cover the reconcile state machines and the CDI provider layer.
  • For production concerns (registry permissions, RBAC, the build-installer bundle), see the project README and the Kubebuilder documentation.