Capacity models
Tenant workloads run on private nodes, machines that join one cluster and nothing else. Because no other tenant's workloads share a private node, its kernel and hardware belong to that tenant alone.
The capacity decision is where those machines come from. A platform admin can reserve machines for a tenant, so its capacity is guaranteed. Or a tenant can draw machines on demand from a pool that no tenant holds. Most offerings combine the two, and for a provider, the choice shapes what you sell, guaranteed hardware or capacity on demand.
Private nodes
A private node joins a single cluster. You can join machines yourself, or let Auto Nodes provision and remove them on demand through a vCluster Platform node provider. See Private nodes and Auto Nodes.
The same machines also back instances with no cluster, which tenants reach over SSH. A machine provisioned as an instance today can join a cluster as a private node tomorrow. See Bare metal as a service.
Reserved capacity
Reserved capacity sets machines aside for one tenant. A platform admin assigns machines to the tenant from the machine inventory with Assign to tenant. From then on, only that tenant's requests can use them.
Use reserved capacity when a customer pays for guaranteed hardware, or when a tenant needs specific machines, such as a GPU class you've committed to it. See Machine inventory.
On-demand capacity
On-demand capacity uses only unassigned machines, those that no tenant owns or holds. Requests never use reserved machines, including the requesting tenant's own machines.
Each node type sets which pool its requests may draw from with a machine allocation policy, or inherits the default from its node provider.
| Policy | UI label | Which machines a tenant's request can use |
|---|---|---|
Assigned | Reserved hosts only | Only machines the requesting tenant owns or that are reserved for it |
OnDemand | Unassigned hosts only | Only machines that no tenant owns or holds |
Metal3 and NVIDIA Infra Controller (NICo) providers apply the policy to the servers they provision. When neither the node type nor the node provider sets a policy, they use Assigned.
KubeVirt ignores the policy and runs every VM on the KubeVirt cluster's own nodes. See Node providers and node types.
Quotas limit capacity, they don't reserve it
Quotas and allow-lists cap what a tenant can consume. None of them sets machines aside, so pair them with reserved capacity when a tenant needs a guarantee.
- Tenant node type quotas cap how many instances of each node type a tenant can hold across all its projects. A platform admin sets them on the tenant. Instances on the tenant's reserved machines don't count, so the quota caps on-demand use. Every KubeVirt instance counts. See Grant object access.
- Project quotas cap what one project can hold. See Quotas.
- Allowed node types limit which node types a project can request. See Allowed node types.
Choose a capacity model
- Guaranteed hardware for a customer. Reserve machines for the customer's tenant, and use the
Assignedpolicy on its node types. - A guaranteed baseline with room to burst. Use separate
AssignedandOnDemandnode types. Reserve machines for theAssignedtype, and cap theOnDemandtype with a tenant node type quota. - Elastic demand from a common pool of machines. Leave machines unassigned, use the
OnDemandpolicy, and give each tenant a quota.
Clusters without private nodes
A cluster created without private nodes runs its workloads on the control plane cluster's own nodes, optionally pinned to a labeled subset of them. Tenants then share those nodes and their kernel, so this model suits trusted internal teams only. See Decide whether tenants share hardware.
Related
- Roles and scopes explains who assigns machines and sets quotas.
- Bare metal as a service and VMs as a service apply these models to instances with no cluster.
- Kubernetes as a service applies them to clusters.