Decommission a Specific Pod in KubeBlocks-Managed Milvus Clusters

This guide explains how to decommission (take offline) specific Pods in Milvus clusters managed by KubeBlocks. Decommissioning provides precise control over cluster resources while maintaining availability. Use this for workload rebalancing, node maintenance, or addressing failures.

Why Decommission Pods with KubeBlocks?

In traditional StatefulSet-based deployments, Kubernetes lacks the ability to decommission specific Pods. StatefulSets ensure the order and identity of Pods, and scaling down always removes the Pod with the highest ordinal number (e.g., scaling down from 3 replicas removes Pod-2 first). This limitation prevents precise control over which Pod to take offline, which can complicate maintenance, workload distribution, or failure handling.

KubeBlocks overcomes this limitation by enabling administrators to decommission specific Pods directly. This fine-grained control ensures high availability and allows better resource management without disrupting the entire cluster.

Prerequisites

Before proceeding, ensure the following:

• Environment Setup:
  • A Kubernetes cluster is up and running.
  • The kubectl CLI tool is configured to communicate with your cluster.
  • KubeBlocks CLI and KubeBlocks Operator are installed. Follow the installation instructions here.
• Namespace Preparation: To keep resources isolated, create a dedicated namespace for this tutorial:

kubectl create ns demo
namespace/demo created

Deploy a Milvus Cluster

Please refer to Deploying a Milvus Cluster with KubeBlocks to deploy a milvus cluster.

Decommission a Pod

Expected Workflow:

1. Replica specified in onlineInstancesToOffline is removed
2. Pod terminates gracefully
3. Cluster transitions from Updating to Running

Before decommissioning a specific pod from a component, make sure this component has more than one replicas. If not, please scale out the componen ahead.

E.g. you can patch the cluster CR with command, to declare there are 3 replicas in component querynode.

kubectl patch cluster milvus-cluster -n demo --type='json' -p='[
  {
    "op": "replace",
    "path": "/spec/componentSpecs/4/replicas",
    "value": 3
  }
]'

To decommission a specific Pod (e.g., 'milvus-cluster-querynode-1'), you can use one of the following methods:

apiVersion: operations.kubeblocks.io/v1alpha1
kind: OpsRequest
metadata:
  name: milvus-cluster-decommission-ops
  namespace: demo
spec:
  clusterName: milvus-cluster
  type: HorizontalScaling
  horizontalScaling:
  - componentName: querynode
    scaleIn:
      onlineInstancesToOffline:
        - 'milvus-cluster-querynode-1'  # Specifies the instance names that need to be taken offline

kubectl get ops milvus-cluster-decommission-ops -n demo -w

NAME                              TYPE                CLUSTER          STATUS    PROGRESS   AGE
milvus-cluster-decommission-ops   HorizontalScaling   milvus-cluster   Running   0/1        8s
milvus-cluster-decommission-ops   HorizontalScaling   milvus-cluster   Running   1/1        31s
milvus-cluster-decommission-ops   HorizontalScaling   milvus-cluster   Succeed   1/1        31s

apiVersion: apps.kubeblocks.io/v1
kind: Cluster
spec:
  componentSpecs:
    - name: milvus
      replicas: 2       # explected replicas after decommission
      offlineInstances:
        - milvus-cluster-querynode-1   # <----- Specify Pod to be decommissioned
 ...

Verify the Decommissioning

After applying the updated configuration, verify the remaining Pods in the cluster:

kubectl get pods -n demo -l app.kubernetes.io/instance=milvus-cluster

Example Output:

NAME                         READY   STATUS    RESTARTS   AGE
milvus-cluster-querynode-0   2/2     Running   0          25m
milvus-cluster-querynode-2   2/2     Running   0          24m

Summary

Key takeaways:

• Traditional StatefulSets lack precise Pod removal control
• KubeBlocks enables targeted Pod decommissioning
• Two implementation methods: OpsRequest or Cluster API

This provides granular cluster management while maintaining availability.