New to KubeDB? Please start here.

Vertical Scaling Milvus

This guide will give an overview on how the KubeDB Ops-manager operator vertically scales a Milvus database.

Before You Begin

• You should be familiar with the following KubeDB concepts:
  • Milvus
  • MilvusOpsRequest

How Vertical Scaling Process Works

Vertical scaling changes the CPU/memory resources of Milvus pods. A MilvusOpsRequest of type VerticalScaling carries the new resources under spec.verticalScaling, keyed by the component you want to scale:

• Standalone: use the node key.
• Distributed: use any of proxy, mixcoord, datanode, querynode, streamingnode (you can scale several at once).

spec:
  type: VerticalScaling
  verticalScaling:
    node:                 # 'node' for standalone; role names for distributed
      resources:
        requests:
          cpu: "1"
          memory: "2Gi"
        limits:
          cpu: "1"
          memory: "2Gi"

The flow is:

1. A user creates a MilvusOpsRequest of type VerticalScaling.
2. The operator validates the request and pauses the Milvus database.
3. The operator updates the resources in the Milvus object’s spec.podTemplate and the corresponding PetSets.
4. Pods are restarted (evicted and recreated) so they come up with the new resources.
5. The operator resumes the database and marks the MilvusOpsRequest as Successful.

Vertical Scaling Modes

KubeDB actuates vertical scaling in one of two modes, selected through the spec.verticalScaling.mode field of the MilvusOpsRequest:

• Restart (default): The operator patches the PetSet with the new resources and restarts the Pods (one at a time, honoring the database’s failover rules) so they come back with the updated CPU and Memory. This works on every Kubernetes cluster.
• InPlace: The operator resizes the running containers in place using the Kubernetes in-place Pod resize (pods/resize subresource) — no Pod restart, so scaling happens without downtime or failover. If a Node cannot accommodate the new resources (the resize is reported Infeasible), the operator automatically falls back to the Restart behavior for that Pod.

If spec.verticalScaling.mode is omitted, it defaults to Restart.

Note: InPlace mode relies on the Kubernetes InPlacePodVerticalScaling feature gate, which is enabled by default from Kubernetes v1.33. On older clusters, or when the feature gate is disabled, use Restart mode.

In the next doc, we will see a step-by-step guide on vertically scaling a Milvus database.