Vector Databases · milvus-io

milvus

Milvus is an open-source, cloud-native vector database designed for large-scale similarity search on unstructured data like text and images. It scales horizontally on Kubernetes, supports multiple index types (HNSW, IVF, DiskANN), and offers both standalone and distributed deployment modes.

Talk to DEV.co Discuss Implementation

Source: GitHub — github.com/milvus-io/milvus

45.1k

GitHub stars

4.1k

Forks

Primary language

Apache-2.0

License (OSI-approved)

Key facts

Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.

Field	Value
Repository	milvus-io/milvus
Owner	milvus-io
Primary language	Go
License	Apache-2.0 — OSI-approved
Stars	45.1k
Forks	4.1k
Open issues	969
Latest release	v2.6.19 (2026-06-26)
Last updated	2026-07-07
Source	https://github.com/milvus-io/milvus

What milvus is

Written in Go and C++ with hardware acceleration (CPU/GPU, NVIDIA CAGRA), Milvus implements a microservices architecture that decouples compute and storage, supports sparse and dense vectors, metadata filtering, and hybrid search. The system includes support for multi-tenancy at database/collection/partition levels and provides TLS encryption and RBAC.

Quickstart

Get the milvus source

Clone the repository and explore it locally.

terminalbash

1git clone https://github.com/milvus-io/milvus.git2cd milvus3# follow the project's README for install & configuration

Need it deployed, integrated, or customized instead? DEV.co ships production installs.

Best use cases

LLM-powered RAG and semantic search applications

Efficiently store and search embedding vectors at scale for retrieval-augmented generation pipelines, full-text search with BM25, and hybrid search combining semantic and keyword matching.

Multi-tenant SaaS platforms with shared infrastructure

Milvus's multi-tenancy support at database, collection, and partition levels enables hundreds to millions of isolated tenants on a single cluster, reducing operational overhead.

Real-time image and content similarity at scale

Support for GPU indexing (CAGRA), distributed architecture, and horizontal scaling enable billions of vector queries for recommendation systems and image search with sub-millisecond latency.

Implementation considerations

Choose deployment mode early: Milvus Lite (Python, local SQLite-like) is suitable for dev/quick prototypes; Standalone (Docker) for single-machine; Distributed (K8s) for production scale. Migration between modes requires planning.
Vector dimensionality and index type selection (HNSW, IVF, FLAT, DiskANN) directly impact memory/CPU and query latency; benchmark with production data and traffic patterns before committing.
Metadata filtering and scalar indexing are required for real-world hybrid queries; default configurations may not suit your cardinality and selectivity patterns.
For multi-tenant deployments, carefully design partition-key strategies and RBAC policies; poor isolation can lead to noisy neighbor effects and security risks.
Persistent storage backend (S3, GCS, Minio, etc.) and message broker (Kafka, Pulsar, RabbitMQ) selection impacts operational complexity, cost, and consistency semantics; test failover scenarios.

When to avoid it — and what to weigh

Simple key-value or relational queries are primary — Milvus is vector-first; if your workload is dominated by SQL joins, transactions, or ACID guarantees, traditional databases are a better fit.
Minimal DevOps expertise and no Kubernetes available — While Milvus Lite and Standalone mode reduce complexity, the distributed mode requires Kubernetes orchestration, etcd, message queues, and operational maturity. Zilliz Cloud exists to mitigate this, but still requires vendor dependency.
Strict data sovereignty or air-gapped environments without managed services — Self-hosted Milvus requires external dependencies (etcd, message brokers, S3-compatible storage). Compliance and security reviews must cover the entire stack.
Extremely low-latency requirements with sub-100ms budgets for all operations — While Milvus is optimized for vector search latency, network and cluster coordination overhead in distributed mode may not meet strict <50ms SLAs for end-to-end requests.

License & commercial use

Milvus is distributed under Apache License 2.0 (Apache-2.0), a permissive OSI-approved license. The project is governed by LF AI & Data Foundation with Zilliz as primary contributor.

Apache 2.0 permits commercial use, modification, and distribution with minimal restrictions (retain license notice and state changes). No explicit patent grants or warranties. For production use, review Apache 2.0 terms and consult legal counsel regarding derivative works and liability. Zilliz offers commercial managed services (Zilliz Cloud) as an alternative to self-hosting.

DEV.co evaluation signals

Editorial assessment — not user reviews. Directional, with an explicit confidence level.

Signal	Assessment
Maintenance	Active
Documentation	Adequate
License clarity	Clear
Deployment complexity	High
DEV.co fit	Strong
Assessment confidence	High

Security considerations

Milvus implements mandatory user authentication, TLS encryption for transport, and role-based access control (RBAC). No security audit or vulnerability disclosure policy mentioned in provided data. Self-hosted deployments must secure all external services (etcd, message broker, S3). Data at rest encryption not explicitly described. Review network segmentation, secret management, and audit logging based on your compliance requirements.

Alternatives to consider

Weaviate

Also open-source, K8s-native vector database with GraphQL API and HNSW indexing. Lighter operational footprint than Milvus for smaller deployments; different schema and query model.

Pinecone

Fully managed, serverless vector database with lower operational burden but vendor lock-in and no on-prem option. Best for teams prioritizing simplicity over control.

Qdrant

Open-source, single-binary vector database with built-in filtering and replication. Simpler deployment than Milvus for moderate scale; less horizontally distributed architecture.

Software development agency

Build on milvus with DEV.co software developers

Start with Milvus Lite for prototyping, benchmark index types and search patterns against production data, and assess Kubernetes and external dependency complexity before committing to distributed deployments. Review Apache 2.0 license terms and security posture for your compliance requirements.

Talk to DEV.co

Related open-source tools

Surfaced by semantic similarity across the DEV.co open-source index.

pymilvus vearch weaviate

Related on DEV.co

Explore the category and the services that help you build with it.

Vector Databases AI Application Development Cloud Deployment API Development Talk to DEV.co

milvus FAQ

Can I use Milvus for production without Kubernetes?

Yes. Milvus Lite works as a Python library for development. Standalone mode (Docker) is suitable for single-machine production with modest throughput. Distributed mode (recommended for scale) requires Kubernetes and external dependencies.

What vector index should I choose?

HNSW is fast and memory-efficient for most use cases. IVF suits larger datasets with fine-tuning. FLAT is exact but slow for billions of vectors. DiskANN optimizes memory/speed trade-off. GPU CAGRA is fastest for NVIDIA hardware. Benchmark with your data.

Does Milvus support transactions or ACID guarantees?

No. Milvus is optimized for vector search, not transactional workloads. It provides eventual consistency in distributed mode and batch insert semantics, but not row-level ACID. Use a relational database if transactions are required.

How much operational overhead is self-hosted Milvus?

Moderate to high. You must provision and monitor Kubernetes, etcd, message brokers, object storage, and Milvus components. Zilliz Cloud eliminates this overhead but introduces vendor dependency. Evaluate your ops maturity and cost tolerance before self-hosting.

Software development & web development with DEV.co

DEV.co helps companies turn open-source tools like milvus into production software. Our software development services cover the full lifecycle — architecture, web development, integration, and maintenance — delivered by software developers and web developers who ship. Engage our software development agency to implement or customize it for your vector databases stack.

Ready to evaluate Milvus for your vector search workload?

Talk to DEV.co Schedule a Consultation