ScalingProfile interface

Scaling configuration for a model deployment. Exactly one of manual or autoscale must be set.

This mutual-exclusion constraint is enforced by the service at request validation time, not by the schema. A PUT request that sets both manual and autoscale, or sets neither, is rejected with HTTP 400 (Bad Request) and an InvalidScalingProfile error code;

Scale-to-zero semantics differ between the two modes:

  • manual permits replicas: 0. This is an explicit operator action to stop serving traffic while keeping the ModelDeployment resource (and its configuration) in place. While at zero replicas the endpoint returns errors for inference requests, and the deployment releases its GPU capacity.
  • autoscale does not permit minReplicas: 0. Autoscaling decisions are driven by serving-server runtime metrics (request rate, queue depth, GPU utilization); at zero replicas there is no signal for the autoscaler to scale back up from. Combined with GPU cold-start time (on the order of minutes) and constrained regional GPU capacity, a scale-from-zero event would produce unacceptable first-request latency and a high risk of capacity unavailability. Callers that want autoscaling with an idle state should delete the ModelDeployment instead.

Properties

autoscale

Autoscaling configuration. Mutually exclusive with manual.

manual

Manual scaling configuration with a fixed replica count. Mutually exclusive with autoscale.

Property Details

autoscale

Autoscaling configuration. Mutually exclusive with manual.

autoscale?: AutoscaleProfile

Property Value

manual

Manual scaling configuration with a fixed replica count. Mutually exclusive with autoscale.

manual?: ManualScalingProfile

Property Value