Operations
Day-2 operations for the Metal3 driver center on managing BareMetalHost resources on the control plane cluster. This page covers adding and removing servers, checking fleet capacity, and where to find guidance for other common tasks. Commands on this page use <control-plane-context>, the cluster referenced by NodeProvider.spec.metal3.clusterRef.cluster.
Set your namespace once. Every command on this page updates automatically.
Add servers​
A physical server joins the pool by creating a BareMetalHost resource and a Secret containing its BMC credentials, both in the NodeProvider's namespace on the control plane cluster. See Add servers for the full field reference and bulk registration.
After creation, the BMO begins registration immediately. The server moves through registering -> inspecting -> available automatically. Inspection collects hardware inventory and typically takes a few minutes. The server isn't eligible for provisioning until it reaches available.
Verify a server is ready​
kubectl --context <control-plane-context> get baremetalhost server-01 -n my-metal3-namespace \
-o jsonpath='{.status.provisioning.state}'
A response of available means the server is ready to claim. See Troubleshooting if the server stalls in registering or inspecting.
Check hardware details​
After inspection, the BareMetalHost records the server's hardware inventory:
kubectl --context <control-plane-context> get baremetalhost server-01 -n my-metal3-namespace \
-o jsonpath='{.status.hardware}' | jq .
This includes CPU count, memory, NIC information, and disk inventory. Verify it matches the expected hardware before the server is provisioned into production.
Inspect the BMC​
Redfish Explorer lets a platform admin inspect a server's BMC resources and run the actions its firmware exposes. Use it to investigate hardware health or BMC behavior.
Before you begin:
- The Metal3 UI requires a platform admin outside tenant scope, with the machine management and Metal3 features enabled. You also need permission to
createthemetal3bmcs/redfishsubresource incluster.loft.sh. The permission applies on the host's control plane cluster, in its namespace. Permission to read a Machine alone doesn't grant BMC access. - The host needs a Redfish-compatible BMC address and a credential Secret. An
ipmi://address doesn't support this explorer. - The platform agent on the control plane cluster must be able to reach the BMC. See BMC requirements.
Redfish access includes actions that change hardware state. A reset or power action can interrupt workloads. Grant this access only to operators trusted to administer the physical server.
Open Bare Metal > Machines and click the Metal3 tab, if shown.
Open the server, then click the Redfish tab. The entry appears only when your permissions allow Redfish requests for that host.
Browse the resource tree or enter a path beginning with
/redfish/. Inspect the resource's properties and links. The BMC's firmware determines which resources appear. A successful request displays the selected resource's properties.
To run an action, click it under Actions. Supply its parameters and any required JSON body, then click . Review the target and request body before clicking .
Read the response before retrying an action. A BMC can accept an asynchronous operation and return a task link. Follow that link to check progress. If a request times out, check the hardware state before sending the action again.
Audit BMC requests​
The platform agent sends requests using the credentials referenced by the BareMetalHost. Requests and responses can contain BMC account passwords and session tokens.
The built-in audit policies record create requests to metal3bmcs/redfish at Metadata level. This preserves attribution without logging request or response bodies. With a custom policy, keep this rule before broader rules that capture bodies. See Metadata-only for sensitive data.
Check a server's address and power​
Open Bare Metal > Machines. The inventory shows On or Off beside each machine's name. On the Metal3 tab, each host's name carries a power badge that also shows transitions, and the IP address column shows the host's configured DHCP address.
Interpret the IP address​
The Metal3 host list shows the address configured in metal3.vcluster.com/ip-address, without its prefix length. It doesn't fall back to an address found during hardware inspection, so the column is empty when the annotation isn't set.
Machine network details in the inventory show inspected NIC addresses, with the configured address on the boot NIC. The Metal3 server details page shows inspected addresses only. Neither view reflects an address set by custom network data, and address information can lag a reset until power-off.
A configured address doesn't mean the server is reachable. Test routing and access from the intended client separately. See Set a host address for configuration and Unexpected or missing IP address when an address differs from what you expect.
Interpret power status​
The Metal3 tab's badge shows On, Off, Powering On, Powering Off, or Rebooting. The transitional labels show that a change is in progress, not that it finished.
Power is separate from provisioning phase. An On server can still be inspecting, provisioning, or failing to boot. During provisioning or deprovisioning, the badge can lag the server's actual power state.
Check the host's Ironic view, when your permissions expose it, to compare current and target power state. See Unexpected power status if the badge disagrees with the server.
Power a claimed server on, off, or reboot it​
To power a server that an instance holds, use the instance's Start, Stop, and Reboot actions or its NodeClaim. The platform keeps the server in the power state that the NodeClaim asks for. A server powered off another way, such as through the BMC, is powered back on while the instance asks for On. Reboot is a hard reboot. See Power control.
Remove servers​
Permanent removal​
If an instance holds the server, delete it from the vCluster Platform UI, or delete its NodeClaim through the API, first. The platform deprovisions the server and returns it to available. Then delete the BareMetalHost.
kubectl --context <control-plane-context> delete baremetalhost server-01 -n my-metal3-namespace
After the normal sequence, the instance and its NodeClaim are already gone and the platform has released the IP address. Deleting the now-unclaimed BareMetalHost only removes the server from inventory.
Also delete its BMC credential Secret if it is no longer needed.
kubectl --context <control-plane-context> delete secret server-01-bmc -n my-metal3-namespace
If the server is permanently unreachable and can't complete normal deprovisioning, delete the BareMetalHost first, then delete its instance. During instance deletion, the driver treats the missing host as already removed, releases the IP, and finishes deleting the NodeClaim. The server doesn't go through the normal deprovisioning cycle, so its disk isn't cleaned. Don't use this sequence for a reachable server that can be deprovisioned normally.
Temporarily withhold a server​
To stop an available server being claimed for new instances, break the selectors that match it. Remove or change a label that every matching node type selector requires. The server stays registered. For example, if the server is eligible because its role is compute, remove that label:
kubectl --context <control-plane-context> label baremetalhost server-01 \
-n <namespace> role-
Merely adding a maintenance=true label doesn't exclude a server unless every applicable node type selector explicitly requires that label not to exist. Label changes don't affect servers that are already provisioned. A server already claimed by an instance remains claimed until that instance is deleted.
Restore the selection label to return the server to the eligible pool:
kubectl --context <control-plane-context> label baremetalhost server-01 \
-n <namespace> role=compute
Check fleet capacity​
Get a view of all servers and their current states:
kubectl --context <control-plane-context> get baremetalhost -n my-metal3-namespace \
-o custom-columns='NAME:.metadata.name,STATE:.status.provisioning.state,ERROR:.status.errorMessage'
Count available servers:
kubectl --context <control-plane-context> get baremetalhost -n my-metal3-namespace -o json | \
jq '[.items[] | select(.status.provisioning.state=="available")] | length'
A server in available is ready for a new instance request. A server in provisioning or provisioned is claimed. A server with operational status error needs attention before it can be used again. See Troubleshooting for how to resolve errors.
Other day-2 tasks​
| Task | Where to find it |
|---|---|
| Release an independently provisioned instance and reclaim its server | Independently provisioned instances |
| Understand what scale-down does to a cluster node | Cluster nodes |
| Update OS images across a fleet | GPU Fleet Operations |
| Update server labels | GPU Fleet Operations |
| Expand the IP address pool | Troubleshooting |
| Rotate BMC credentials | Production and security |