Skip to content

Scale custom targets and prewarm node capacity

Capacity buffers request spare resources before application pods need them. They reduce provisioning delay when the workload fits the ready buffer, at the cost of holding unused capacity.

A demand signal changes buffer chunks through ScaleAdapter; the provider fits those chunks onto suitable ready nodes.
Scroll to explore
Diagram description

A KEDA ScaledObject drives a ScaleAdapter. The adapter writes CapacityBuffer replicas, representing chunks of a pod resource shape. The provider evaluates placement and provisions missing node capacity. Real application pods can then use suitable Ready nodes. Buffer chunks are capacity requests, not running application pods, and chunk count is not node count.

ScaleAdapter exposes a custom resource through Kubernetes’ scale interface. For a CapacityBuffer, it changes the requested number of pod-shaped chunks, not the number of nodes. The provider decides how those chunks fit existing or new nodes. An adapter’s desired count does not prove that nodes or application pods are Ready.

Give the target field one writer. See ScaleAdapter API reference for other targets, including Agones Fleets.

EnvironmentSetupVerify readinessConstraints
KarpenterCreate and scale a bufferBuffer conditions and NodeClaimsNodePool limits, placement and disruption settings.
Cluster AutoscalerCreate and scale a bufferBuffer conditions and nodesSupported controller versions, node-group limits and scheduling constraints.
GKECreate and scale a bufferBuffer conditions and nodesStandby buffers exclude GPU/TPU nodes; active buffers have separate prerequisites.

Use the provider’s verification commands above. Confirm that Ready nodes have the resources, zones and taints/affinity required by the application; requested buffer chunks alone are insufficient. Measure provisioning plus application startup time to set schedule or forecast lead time.

When the buffer shrinks, node removal follows the provider’s consolidation rules and workload placement. Reducing the buffer does not guarantee immediate cost reduction. For desired replicas without ready pods, use workload diagnostics.