3 Release 2.3.0 #
This release builds on the 2.2.x line with customizable blueprint deployments, custom chart repositories and Git-sourced blueprint catalogs, clearer repository and workload health reporting, and a new workload details view, plus several stability and installation improvements.
3.1 What’s new #
3.1.1 Blueprint customization #
Customize blueprint deployments: A new
Configurationstep in the blueprint installation and management wizard lets you override Helm values and enable or disable individual blueprint components before deployment. Disabled components are left out of the deployment and removed from the cluster. Blueprint workloads can be reconfigured later with the action #265.Configuration drift tracking: Workloads report whether their configuration differs from the blueprint defaults, and the
Workloadspage shows aCustomizedbadge next to the blueprint source. Only the values you changed are saved, not the full set of component values #287.Partner icons and
Partnersource: Blueprints accept an optionalicon, either anhttps://URL or a base64-encoded rasterdata:URI of up to 16 KB, which is shown next to the source badge, and a newPartnersource value. Updating a blueprint with an invalid specification now returns HTTP422instead of500#292.
3.1.2 Custom blueprint catalogs #
Git-sourced blueprint catalogs: Administrators can add external Git repositories that contain blueprint definitions in the
Custom Catalogssection of Settings or through theblueprintCatalogsHelm value. The operator synchronizes each catalog into the cluster through its own FleetGitReporesource, with its own credential and CA certificate secrets #281.New
BlueprintCatalogCRD: A cluster-scoped resource that describes a curated, branded collection of blueprints #281.Browse blueprints by catalog: The
Blueprintspage provides a catalog filter. Blueprints bundled with the chart are listed underSUSE Blueprints#281.
3.1.3 Custom chart repositories #
Manage chart repositories from Settings: Administrators can add, edit and delete Helm (HTTP or HTTPS), OCI and Git chart repositories, including credentials, in the
Repositoriessection of Settings. The operator creates aClusterReporesource with the same name for each repository, and its applications are listed in the application catalog under their own library #244.Single-chart OCI repositories: Repositories whose URL already ends with the chart name, for example
oci://ghcr.io/nvidia/openshell/helm-chart, install correctly with both Helm and Fleet. Use this to add the NVIDIA OpenShell charts #244, #293.Verify before saving: Custom repositories can be tested for credentials, reachability and access to a sample chart from the same form #244.
3.1.4 Repository verification and readiness #
Three-part connection test: In Settings, reports the registry connection, access to a sample chart, and the readiness of the Rancher Prime chart repository as separate results, so a successful login alone is no longer reported as success. The result names the chart and version that were tested, and denied access, missing content, invalid indexes, TLS errors and timeouts each have their own explanation #242.
Repository failures shown across the UI: The
Overviewpage lists failing managed repositories with links to Settings and Rancher Prime. Applications from an unavailable repository remain visible but cannot be installed, and the installation wizard explains missing repository indexes instead of reportingconfigmaps "" not found#242.Explicit refresh: is read-only. A new action asks Rancher Prime to download the index of a saved repository again #242.
3.1.5 Workload monitoring #
Workload details: Selecting a workload name on the
Workloadspage opens a details panel with the source, deployment strategy, target namespace and clusters, blueprint components, per-cluster status messages, value overrides, and the underlying Fleet resources or Helm release #221, #303.Accurate status for Helm workloads: Application workloads deployed with Helm now derive their status from pod and controller readiness. A crash-looping or unschedulable pod reports
Degradedwith the reason, and a pod that is still starting reportsPendingfor as long as it needs. Previously, a workload was reported asRunningas soon as its Helm release existed #260.
3.1.6 Air-gapped deployments #
NVIDIA repository aliases preserved in mirror mode: Switching to a private NVIDIA mirror keeps every repository alias derived from the catalog, such as
nvidia-runaiornim-nvidia, so existing workloads can still be upgraded and redeployed #237.Bundled catalog logos: Logos for 81 SUSE AI charts are bundled with the UI extension, so application tiles show the correct logo in connected and air-gapped installations, including behind private mirrors #245.
3.1.7 Application catalog #
NVIDIA catalog refresh: The weekly NVIDIA AI Enterprise catalog refreshes are included. NGC applications that lost their NVIDIA AI Enterprise support label, including several NeMo microservices, two NIM embedding models and
dgxc-admission-controller, remain in the catalog without theSupportedlabel. The NVIDIA GPU Operator no longer has theSupportedlabel #301.
3.1.8 Packaging #
3.2 Key fixes #
The UI extension no longer remains in the
Failedstate after the operator re-creates itsaif-ui-configConfigMap. The operator now takes Helm ownership of an existing ConfigMap instead of failing every installation and upgrade with aninvalid ownership metadataerror #218.Custom UI extension settings are retained after uninstalling. The
aif-ui-configConfigMap was previously deleted with the release, so custom operator connection and catalog settings were lost on reinstallation #218.Failing Helm workloads are no longer reported as
Running. For example, a NIM pod that did not fit in GPU memory was reported asRunningwhile it was crash-looping #260.Kubeflow, LiteLLM and Qdrant no longer fail on the first installation wizard screen when chart access is denied or the Rancher Prime repository is disconnected despite a successful registry login #242.
Applications such as Ollama show their own logo instead of a generic icon when their metadata contains a logo URL #245.
Switching to an NVIDIA mirror no longer breaks later upgrades of applications installed from NVIDIA team repositories #237.
3.3 Upgrade notes #
Mirror the
kubectlimage for air-gapped installations: The new Rancher Prime readiness hook job uses theregistry.suse.com/suse/kubectl:1.35image, which you can configure with theglobal.kubectl.imagevalues. The job waits until Rancher Prime is ready, limited only by the installer timeout, such ashelm --timeoutor the Fleethelm.timeoutSecondsvalue #277, #289.New CRD and settings: The
blueprintcatalogs.ai-factory.suse.comCRD and theblueprintCatalogsandcustomRepossettings are added, and theAIWorkloadandBlueprintCRDs are updated. The CRD hook job of the chart applies them on upgrade #281, #244, #265.Use matching operator and UI extension versions: Chart access checks, blueprint customization and partner icons require both the 2.3.0 operator and the 2.3.0 UI extension #242, #265, #292.
Helm workload status may change: Application workloads that were previously reported as
Runningmay now reportPendingorDegradedif their pods are not ready #260.aif-ui-configis retained on uninstall: Uninstalling the UI extension now leaves theaif-ui-configConfigMap in place. Delete it manually for a clean reinstallation #218.Standalone installations of the UI extension chart: Installing or upgrading the
aif-uichart withhelmwithout the operator, over an existingaif-ui-configConfigMap, requires the--take-ownershipoption, available in Helm 3.15 and later #218.Reserved custom repository names: Custom repositories cannot use the names of the built-in SUSE AI Factory repositories, the default Rancher Prime repositories, the NVIDIA team repositories, or a name already used by another
ClusterReporesource #244.Partnerblueprint source: SUSE AI Factory Operator 2.2.0 rejects blueprints with thePartnersource. Do not use it until 2.3.0 is the minimum operator version in your environment #292.
3.4 Known issues and limitations #
Air-gap support covers private chart repositories, private Git repositories, repository credentials and CA trust. Preparing the cluster remains a separate task: node and containerd registry mirrors, DNS, host proxies and operating system repositories are not configured by SUSE AI Factory.
Container image references inside blueprint values are not rewritten. Your containerd registry mirrors must resolve them transparently.
Git over SSH is not supported for blueprint input and GitOps output. Use HTTPS with credentials and, where required, a private CA certificate.
Blueprint partner icons specified as
https://URLs do not load in air-gapped installations. Use base64-encodeddata:URIs instead.A custom HTTPS repository that uses a certificate issued by a private CA is not trusted through the CA bundle alone. As a workaround, enable
Skip TLS verificationfor that repository.Anonymous connection tests against OCI registries that require repository-scoped tokens, such as GitHub Container Registry, report an authentication failure even though the charts can be pulled anonymously.
Chart access is only checked for
https://Helm andoci://repositories. A plainhttp://Helm repository can pass the connection test but reports a configuration error for chart access.Running chart access checks repeatedly against repositories with very large indexes can exceed the operator memory limit and restart the operator.
The OpenShell gateway chart requires the Agent Sandbox CRDs and controller (
agents.x-k8s.io) on the target cluster.A Helm workload whose pod is stuck for a reason that is only reported as a pod event, such as
FailedMountfor a missing secret or ConfigMap, remainsPendinginstead ofDegraded.Upgrade, rollback and retry apply to blueprint workloads only. A workload that has never reached a healthy, fully rolled out state has no rollback target.
Full Changelog: aif-operator-2.2.0…aif-operator-2.3.0