Skip to main content
Version: v1.4

Just-in-Time Workload Clusters

Nova can dynamically manage workload clusters to reduce cost and provide capacity when it is needed.

This includes placing idle clusters into standby, resuming clusters when required, and optionally creating new clusters from existing workload clusters.

When to Use This​

Use this pattern when:

  • Workload demand is variable
  • Idle cloud clusters should be minimized to reduce cost
  • Additional workload cluster capacity may be needed on demand
  • Clusters should be made available only when scheduling requires them

How Nova Helps​

Nova can align workload cluster availability with scheduling demand.

Depending on configuration, Nova can:

  • Put idle workload clusters into standby
  • Resume standby clusters when needed
  • Delete and recreate clusters
  • Create new clusters by cloning existing workload clusters

Standby Modes​

Suspend/Resume​

In suspend/resume mode, Nova scales cluster node pools to zero while keeping the cluster control plane intact.

  • Faster recovery (~2 minutes)
  • Lower cost savings compared to full deletion

Delete/Recreate​

In delete/recreate mode, Nova deletes the cluster entirely and recreates it when needed.

  • Greater cost savings
  • Longer startup time (~3–15 minutes)

Cluster Creation​

When enabled, Nova can create new workload clusters by cloning an existing cluster.

This is used when:

  • No existing cluster satisfies placement requirements
  • Additional capacity is required
  • No existing workload cluster is running an autoscaler recognized by Nova (such as Elotl Luna or Kubernetes Cluster Autoscaler) that can provision the required capacity

Cluster creation is driven by the source cluster configuration. When an autoscaler is available, Nova will prefer to use it for provisioning and scaling before creating a new workload cluster.

Behavior with Autoscaling​

When an autoscaler is available, Nova will leverage it for node provisioning and scaling, and will prefer using the autoscaler over just-in-time cluster creation when sufficient capacity can be provisioned this way. This includes integrations with autoscalers such as Elotl Luna and Kubernetes Cluster Autoscaler, though their presence is optional and not required for just-in-time provisioning.

Considerations​

Just-in-time cluster behavior depends on:

  • Cloud provider support
  • Cluster creation and startup latency
  • Nova configuration
  • Whether candidate clusters already have autoscaling enabled
  • Whether an existing cluster is available to clone