PROVIDER Β· Open source distributed vector database
Milvus
Milvus is the distributed vector database open sourced by Zilliz and donated to LF AI, where multiple index types and separated storage and compute balance performance and cost at billion-vector scale as the mainstream self-hosted RAG choice.
Pros
Cons
Product Information
Detailed Notes
Overview
Milvus was open sourced by Zilliz in 2019 and donated to the LF AI & Data Foundation, written in Go and C++ with separated storage and compute plus a distributed architecture coordinating etcd, object storage, message queues, and query nodes, the most representative open source vector database.
For Backend & Database work the value is controllable cost at large vector scale, since HNSW, IVF, DiskANN, and GPU indexes can be chosen for data size and latency needs, and the distributed architecture holds billions of vectors while self-hosting avoids long-term usage billing.
The limit is operational complexity, because a distributed cluster carries many components including etcd, MinIO or S3, Pulsar or Kafka, and query, index, and data nodes, costing far more effort than managed options, and small deployments should use Lite or Standalone while noting capability differences from the cluster edition.
Core Strengths
- Multiple index types including HNSW, IVF, DiskANN, and GPU for size and latency trade-offs
- Separated storage and compute scale horizontally to billions of vectors
- Hybrid scalar and vector filtering satisfies metadata constraints in RAG
- Apache 2.0 under LF AI governance with no vendor lock-in
Product Ecosystem
Indexing and retrieval
HNSW targets low-latency recall, IVF variants suit memory-constrained libraries, and DiskANN enables disk-level indexes to cut cost.
Distributed components
etcd stores metadata, object storage holds data, message queues carry logs, and query plus index nodes scale independently.
Ecosystem and tooling
The Attu admin UI, multi-language SDKs, and native integration with LangChain and LlamaIndex.
Drawbacks
- Many distributed components raise the operations and tuning bar above managed stores
- Wrong memory and index parameters such as nlist, M, and ef visibly hurt recall
- Lite and cluster editions differ, forcing index strategy review on migration
- English community discussion trails the pgvector ecosystem
Use Cases
- RAG knowledge base retrieval (β β β β β ) - Hybrid filtering with high recall
- Billion-scale vector search (β β β β β ) - Distributed horizontal scaling
- Image and multimodal search (β β β β ) - Multiple indexes with GPU acceleration
- Small-scale validation (β β β β ) - Start with Lite or a managed service
Pricing
| Plan | Price | Scope |
|---|---|---|
| Milvus open source | $0 | Apache 2.0, self-managed |
| Zilliz Cloud | Per compute unit and storage | Managed with on-demand scaling |
| Other managed offerings | Per instance size | Self or semi-managed deployment |
The open source edition is free while a cluster needs several machines plus object storage and a message queue, with standalone lowering the bar for small scale, and managed billing per compute unit wins below tens of millions of vectors while self-hosting pays off beyond hundreds of millions.
FAQ
-
Milvus or pgvector? pgvector is simplest within tens of millions of vectors on existing PostgreSQL, while Milvus fits hundreds of millions or when multiple index types and distributed scale are needed.
-
How is poor recall fixed? Check index type and parameters such as HNSW M and ef or IVF nlist and nprobe, then whether the embedding model suits the language and domain, since parameters and model matter more than the database.
-
Can it run standalone? Yes with Lite and Standalone for development and mid-scale, while production HA needs the full distributed deployment with different capabilities and operations.