How do I deploy Milvus on Kubernetes using KubeBlocks?

Install KubeBlocks via Helm, then install the Milvus addon: "helm upgrade -i kb-addon-milvus kubeblocks/milvus -n kb-system". Create a cluster by applying a Cluster resource with topology: standalone or topology: cluster. KubeBlocks provisions all required pods, services, and storage automatically.

What is the difference between Milvus Standalone and Distributed topology in KubeBlocks?

Standalone runs all Milvus coordinator and worker roles as goroutines inside a single pod — ideal for development, CI, and single-node deployments. Distributed (cluster) runs each component (proxy, mixcoord, datanode, indexnode, querynode) as separate scalable pods backed by MinIO, etcd, and Pulsar — designed for production and billion-scale workloads. Both topologies use the same Cluster CRD API; switch by changing the topology field.

What Milvus versions does KubeBlocks support?

KubeBlocks supports Milvus 2.3.2 and 2.5.13. Rolling version upgrades from 2.3.x to 2.5.x are supported with component-ordered sequencing: worker nodes (datanode, indexnode, querynode) are upgraded first, then proxy, then mixcoord.

Can I scale Milvus horizontally with KubeBlocks?

Yes. KubeBlocks supports independent horizontal scaling of querynode (for more ANN search throughput), datanode (for higher insert and flush concurrency), indexnode (for faster index builds), and proxy (for higher client concurrency). All Milvus workers are stateless — scale out without data migration. Brief segment rebalancing occurs during querynode scale-out.

Does KubeBlocks support Milvus backup and restore?

Yes. KubeBlocks integrates milvus-backup (v0.5.9+) for collection-level snapshots. Backups are exported to S3-compatible object storage (MinIO or any compatible endpoint) and can restore to a new Milvus cluster at any time. Standalone clusters also support volume snapshot backup.

How does Milvus store data when managed by KubeBlocks?

All Milvus compute pods (proxy, mixcoord, datanode, indexnode, querynode) are stateless and use emptyDir volumes. Persistent state is stored externally: vectors and indexes in MinIO (S3-compatible object storage), cluster metadata and collection schemas in etcd, and the write-ahead log in Pulsar or Kafka. This separation of compute and storage allows any worker pod to be scaled or restarted without data loss.

Is the KubeBlocks Milvus Operator open source?

Yes. KubeBlocks is open source under the AGPL-3.0 license at https://github.com/apecloud/kubeblocks. The Milvus addon is available at https://github.com/apecloud/kubeblocks-addons. An Enterprise edition adds a web management UI, cross-cluster standby, and dedicated support.

The Production-Grade Milvus Operator for Kubernetes | KubeBlocks

Open Source · Production-Grade · CNCF Graduated

KubeBlocks Milvus Operator for Kubernetes

Deploy production-grade Milvus vector database clusters in minutes. Standalone and distributed topologies, billion-scale ANN search, and full Day-2 operations.

Try Playground Free →Read the Docs

1B+

Vectors at Scale

< 1ms

ANN Search Latency

Deploy Milvus in 4 steps

Install KubeBlocks

# Add Helm repo
helm repo add kubeblocks https://apecloud.github.io/helm-charts
helm repo update

# Install KubeBlocks
helm install kubeblocks kubeblocks/kubeblocks \
  --namespace kb-system --create-namespace

Install Milvus Addon

helm upgrade -i kb-addon-milvus kubeblocks/milvus \
  -n kb-system

Create a Milvus Cluster

apiVersion: apps.kubeblocks.io/v1
kind: Cluster
metadata:
  name: milvus-cluster
  namespace: demo
spec:
  terminationPolicy: Delete
  topology: standalone
  componentSpecs:
    - name: milvus
      serviceVersion: "2.5.13"
      replicas: 1

Cluster is Ready

$ kubectl get cluster milvus-cluster -n demo
NAME                CLUSTER-DEF   STATUS    AGE
milvus-cluster               Running   2m

Trusted by Engineering Teams at Scale

BONC Cloud

China Mobile Cloud

China Telecom Cloud

Tencent

Xiaomi

Ping An

VIP.com

Kwai

Tiger Brokers

CITIC Securities

SealOS

FastGPT

KubeSphere

Momenta

360

Tilaa

Olares

Changan Automobile

State Grid

BONC Cloud

China Mobile Cloud

China Telecom Cloud

Tencent

Xiaomi

Ping An

VIP.com

Kwai

Tiger Brokers

CITIC Securities

SealOS

FastGPT

KubeSphere

Momenta

360

Tilaa

Olares

Changan Automobile

State Grid

Architecture

From Dev to Production. Same CRD API.

A single Milvus pod runs all coordinator and worker roles as goroutines within one process. etcd and MinIO are deployed as companion pods. Ideal for development, CI pipelines, and single-node deployments.

✓

Single pod runs all roles: Proxy, RootCoord, QueryCoord, DataCoord, IndexCoord, QueryNode, DataNode, IndexNode

✓

etcd and MinIO deployed as companion pods for metadata and object storage

✓

PVC-backed local segment cache for fast cold-start

✓

Minimal resource footprint — runs on a single Kubernetes node

✓

Same CRD API as distributed mode — switch topologies without application changes

All-in-One Pod

Dev & Test

Simple Setup

Application / Client

Milvus SDK (Python / Go / Java / Node) · gRPC or REST
Connect to {cluster}-milvus:19530

gRPC :19530 / metrics :9091

{cluster}-milvus

ClusterIP · :19530 (gRPC) · :9091 (metrics/health)
selector: all standalone pods

ClusterIP

milvus-standalone-0All-in-One

Container

milvus:19530 (gRPC) · :9091 (metrics)

All coordinator + worker roles run as goroutines within the single milvus process

Embedded Roles

RootCoord

QueryCoord

DataCoord

IndexCoord

QueryNode

DataNode

IndexNode

Proxy

PVC data-0 · local segment cache · 20Gi

metadata + object storage I/O

etcd-0

metadata · collection schemas · segment info

etcd:2379

PVC · 10Gi

minio-0

object storage · vectors · indexes · WAL

minio:9000

PVC · 20Gi

Standalone topologyall roles in one process · ideal for development and single-node deployments

Milvus (all-in-one)

Storage Components (etcd / MinIO)

Persistent Storage

Capabilities

Built for Production Milvus

KubeBlocks automates the hardest parts of running Milvus on Kubernetes — so your team doesn't have to.

Backup & Restore

Collection-level snapshot backups via milvus-backup. Export to S3-compatible object storage and restore to a new cluster at any time.

milvus-backup

Backup Timeline

Snapshot

00:00

Snapshot

08:00

Snapshot

16:00

Restore Target

20:17

1. Export Collectionsmilvus-backup exports collection data and indexes to S3

→

2. Upload to S3Backup stored in S3-compatible object storage

→

3. Restore ClusterNew Milvus cluster restored from any stored snapshot

✓ S3-Compatible Storage

Day-2 Operations

Every Operation Declared as a Kubernetes Resource

No SSH into pods, no shell scripts. Submit an OpsRequest and KubeBlocks handles the rest.

Scaling & Availability

✓

Horizontal Scaling

Scale querynode, datanode, indexnode, and proxy independently. All workers are stateless — scale out without data migration.

✓

Vertical Scaling

Resize CPU and memory for any Milvus component via OpsRequest with a rolling restart that keeps the cluster available.

✓

Volume Expansion

Expand PVCs for components that use persistent storage without pod restarts on supported storage classes.

✓

Rolling Restart

Controlled rolling restarts with health checks between each pod to maintain service availability.

✓

Stop / Start

Suspend the Milvus cluster to eliminate compute cost; resume with full state from MinIO and etcd.

Configuration & Data

✓

Dynamic Configuration

Tune Milvus parameters via OpsRequest — dataCoord settings, queryNode cache sizes, etcd endpoints — without full restarts.

✓

Version Upgrade

Rolling upgrades from 2.3.x to 2.5.x with component-ordered sequencing: workers first, proxy, then mixcoord.

✓

Prometheus Metrics

Per-component metrics exposed on port 9091; compatible with Prometheus and Grafana dashboards.

✓

Backup & Restore

Collection-level backup via milvus-backup (v0.5.9+); restore to a new cluster from any stored snapshot.

✓

Expose via LoadBalancer

Expose the gRPC proxy endpoint on port 19530 via a Kubernetes LoadBalancer service for external client access.

From the Blog

Go Deeper on Milvus on Kubernetes

KubeBlocks x Milvus: Kubernetes Operator Guide

Deploy standalone and cluster modes, integrate etcd/MinIO via Addons, and manage Day-2 operations with an operator-first approach.

We Let an AI Agent Manage Our Databases

We tested AI agents against traditional Kubernetes database operators — and why a unified API like KubeBlocks changes everything.

Deploy Production-Ready AIGC Applications with KubeBlocks and Dify

Build and deploy production-grade AIGC applications on Kubernetes with KubeBlocks managing PostgreSQL, Redis, and Qdrant.

Also on KubeBlocks

Apache Kafka Operator for Kubernetes →

Milvus 2.5+ supports Apache Kafka as its write-ahead log (WAL) for higher throughput streaming. KubeBlocks manages both Milvus and Kafka clusters — deploy the full AI data stack on a single operator.

Get Started

Get Started with KubeBlocks Milvus Operator, Risk-Free.

Open source and production-ready. Enterprise customers get dedicated onboarding, migration support, and SLA guarantees.

Try Playground Free →Talk to the Team

✓Open Source

✓Milvus & 35+ other engines

✓Production-grade HA

✓No vendor lock-in

# Add Helm repo helm repo add kubeblocks https://apecloud.github.io/helm-charts helm repo update # Install KubeBlocks helm install kubeblocks kubeblocks/kubeblocks \ --namespace kb-system --create-namespace

apiVersion: apps.kubeblocks.io/v1 kind: Cluster metadata: name: milvus-cluster namespace: demo spec: terminationPolicy: Delete topology: standalone componentSpecs: - name: milvus serviceVersion: "2.5.13" replicas: 1