v1.0.1¶
5th of November, 2026
kMetal v1.0.1 is a patch release on top of v1.0.0. It fixes issues found since General Availability in the installation order, the placement of the networking components, the image builder, and the Operator Console. It also adds support for Kubernetes v1.37 Tenant Control Planes through Kamaji, and lets OS images be pushed to an OCI registry running inside the under cluster. Read the Upgrade Notes before upgrading: the node label that places Kube-OVN has changed.
Features¶
N.R.
Improvements¶
-
feat: node labelling controller
The operator runs a new controller that adds the labels the networking components expect to the under cluster's Nodes, so they no longer have to be applied by hand. Nodes labelled
node-role.kubernetes.io/networkgetovn.kubernetes.io/external-gw=true, and the nodes listed inspec.networking.kubeOVN.centralNodesgetkube-ovn/role=master. A Node is patched only when its labels actually change, so the constant status updates from the kubelet put no extra writes on the API server. -
feat: bump kamaji capi provider to 1.0.2
The Cluster API Control Plane provider for Kamaji moves from v1.0.1 to v1.0.2.
KamajiControlPlanegainsallocateLoadBalancerNodePorts: set tofalse, it stops Kubernetes from allocating NodePorts for a Tenant Control Plane exposed as aLoadBalancerService. The Konnectivity agent's resources can now be set throughspec.addons.konnectivity.agent.resources. -
feat(deps): bumping up cloud-clusterclass-kmetal to v1.10.4
The kMetal ClusterClass moves from v1.10.1 to v1.10.4. Passthrough GPUs are attached with the virtual display disabled: KubeVirt attaches a virtual display adapter by default, and that stops a VM from starting on any headless datacenter GPU, such as the A100 or H100.
KamajiControlPlaneTemplatemoves tov1alpha2, which the Cluster APIv1beta2contract requires. Tenant Control Planes exposed asLoadBalancertake their address from thetenant-cp-poolMetalLB pool. -
feat(deps): bumping up cloud-clusterclass-kmetal to v1.10.5
The kMetal ClusterClass moves to v1.10.5. Tenant Control Plane pods are scheduled on the under cluster's control-plane nodes, tolerating their taint, and each Tenant Control Plane's replicas are spread across those nodes on a best-effort basis. The MetalLB address pool and the address requested through
controlPlane.network.serviceAddressare now set in a single patch, so one no longer overwrites the other. -
feat(image-builder): replicating snapshot size in status
The image builder moves from v1.3.0 to v1.3.1. Some CSI drivers leave
status.restoreSizeunset when they create a snapshot, and CDI refuses to clone from aVolumeSnapshotwithout one. The image builder now copies the source PersistentVolumeClaim's capacity into both the goldenVolumeSnapshotand itsVolumeSnapshotContent, and waits for the snapshot to be ready before cloning from it. -
feat: supporting OCI artefacts from local registry
OS images built with the
ocistrategy can be pushed to a registry running inside the under cluster, such as Harbor, even when its certificate is signed by a private CA. The newimage-builder.clastix.io/registryCASecretannotation names the Secret holding the CA in itsca.crtkey, for example one issued by cert-manager. The ClusterClass exposes it asoci.registryCASecret, next to the newoci.registrySecretfor the push credentials. The newosArtifact.nodeSelectorClusterClass value places the Pods, Jobs, and installer VirtualMachine that build an OS image. This release ships the kMetal ClusterClass v1.10.7, the image builder v1.3.2, and kMetal KubeVirt v1.2.1, which moves KubeVirt from v1.8.2 to v1.8.4 to fix checksum failures with OCI images. -
feat(deps): supporting Kubernetes v1.37 via Kamaji
Kamaji and its CRDs move from 1.4.2 to 1.5.0, adding support for Kubernetes v1.37 Tenant Control Planes.
-
feat: install kube-ovn from clastix fork
Kube-OVN moves from v1.15.5 to v1.15.25 and is now installed from the Clastix fork: the chart comes from the Clastix OCI repository, and the images from
ghcr.io/clastix. The Secret named byspec.registry.imagePullSecretNameis passed to every Kube-OVN pod, so the images can be pulled from the private registry. The upstreamkubeovnHelm repository is no longer added to the under cluster.
Bug Fixes¶
-
fix: replicating image pull secret in Flux's namespace
The image pull secret is now copied into Flux's namespace as well. Every
HelmRepositoryandOCIRepositoryis created there, and Flux looks up a source's credentials in the source's own namespace. Without the copy, the private Clastix sources were fetched anonymously and thekmetal-*charts failed to install, even though every public source reconciled. -
fix(order): image-builder is dependency of clusterclass
The image builder is now installed before the kMetal ClusterClass, which depends on it.
-
fix(image-builder): creating Namespace in advance to accomodate Certificate
The operator now creates the Operator Console's namespace,
kmetal-console, instead of leaving it to the Helm chart. Withspec.console.tlsset, the console's Certificate was rejected because its namespace did not exist yet, and the chart could never finish installing without the Secret that Certificate issues. The Certificate is now applied before the Helm release, so the installation no longer deadlocks. -
fix(kube-ovn): using well-known labels for network component
Kube-OVN's per-node DaemonSets (
ovs-ovn,kube-ovn-cni, andkube-ovn-pinger) are now placed on nodes labellednode-role.kubernetes.io/network, instead of the kMetal-specifickmetal.io/ovn. Nodes no longer need an extra label on top of the kMetal node roles. See the Upgrade Notes. -
fix(kube-ovn): using empty value for node selector
The Kube-OVN placement selects
node-role.kubernetes.io/networkwith an empty value, following thenode-role.kubernetes.io/*convention, instead of"true". -
fix: kmetal-networking has capi dependency
kMetal Networking and the provider networks are now installed after the Cluster API core provider. kMetal Networking indexes Cluster API
Clusterobjects at startup, so it needs their CRD in place first. -
fix(metallb): speaker must run only on CP nodes
The MetalLB speaker now follows
spec.placement, like the controller, instead of running on every node: by default, that means the under cluster's control-plane nodes. Tenant Control Plane addresses are announced from those nodes, so a speaker anywhere else served no purpose. -
fix(capi): fixing control plane placement and service annotations
The kMetal ClusterClass moves to v1.10.6. The
metallb.io/loadBalancerIPsannotation is set on a Tenant Control Plane's Service only whencontrolPlane.network.serviceAddressis set; before, Cluster API rendered<no value>into it. The topology spread constraint added in v1.10.5 no longer lists the same label key in bothlabelSelectorandmatchLabelKeys, which made the API server reject the Tenant Control Plane pods. -
fix(console): supporting OCI for wizard section
The Operator Console plugin moves from v1.0.0-rc1 to v1.0.1. On platforms using the
ociroot-disk strategy, the cluster creation wizard reported "The platform publishes no machine image" and blocked cluster creation. It now reads the available Kubernetes versions fromosArtifact.tags, as it already did for thedataVolumestrategy. -
fix(image-builder): switching over crane for OCI push
The image builder moves to v1.3.3, and now pushes OS images to OCI registries with
craneinstead ofbuildah. The push stores only the compressed disk layer instead of two full copies of the disk, so it needs a much smaller scratch volume. The layer is written non-sparse, since containerd fails to unpack sparse disk images with "sparse map too long", and the image is tagged with the architecture of the node that pushes it. -
fix(kamaji): setting no limits to manager
The Kamaji manager now runs without CPU and memory limits, keeping the chart's requests. The chart's default limits (200m CPU and 100Mi memory) are sized for a handful of tenants. The manager caches every Tenant Control Plane and the objects it owns, so as tenants were added it reached its memory limit and was
OOMKilledin a restart loop, leaving every Tenant Control Plane unreconciled. The CPU limit also throttled the reconciles that would have caught up. -
fix: removing exclude from loadbalancer label from control plane nodes
The node labelling controller removes
node.kubernetes.io/exclude-from-external-load-balancersfrom the under cluster's control-plane nodes, where installers such as kubeadm set it by default. MetalLB does not announce addresses from nodes with that label. Since the speaker now runs only on control-plane nodes, keeping the label would leave no node to announce Tenant Control Plane addresses.
Known Issues¶
N.R.
Upgrade Notes¶
-
Label the Kube-OVN nodes with
node-role.kubernetes.io/networkbefore upgrading.Kube-OVN's per-node DaemonSets are no longer placed by
kmetal.io/ovn. Before upgrading, label every node that carries the overlay withnode-role.kubernetes.io/networkand an empty value. If you skip this, Kube-OVN's DaemonSets are removed from those nodes when the operator upgrades.The operator also labels these nodes with
ovn.kubernetes.io/external-gw=true.kmetal.io/ovnis no longer read, so you can remove it after the upgrade. See Networking Configuration. -
MetalLB announces only from the nodes selected by
spec.placement.The speaker no longer runs on every node: make sure the nodes selected by
spec.placement, the control-plane nodes by default, are attached to the network segment of every MetalLB address pool. The operator removesnode.kubernetes.io/exclude-from-external-load-balancersfrom control-plane nodes, and removes it again if anything sets it back. -
Kube-OVN images now come from
ghcr.io/clastix.Kube-OVN is upgraded from v1.15.5 to v1.15.25, and its DaemonSets roll to the Clastix images during the upgrade. These images are pulled with the Secret named by
spec.registry.imagePullSecretName, the same one the kMetal charts already need.