Milvus
Cloud-native vector database for billion-scale enterprise AI
Our Take
Milvus is the most mature and feature-rich open-source vector database. Its C++/Go core with a 4-layer disaggregated architecture (access → coordinator → worker → storage) enables true billion-scale with independent scaling. The only major vector DB with full GPU acceleration via NVIDIA CUDA (CAGRA, GPU_IVF). Widest index selection: 12+ types including HNSW, IVF_FLAT, IVF_PQ, DiskANN, and SCANN. Won BigANN at NeurIPS 2021. VDBBench shows Zilliz Cloud at 9,704 QPS with 2.5ms p99 — top of the leaderboard. ~43,000 GitHub stars, 300+ contributors. The tradeoff: cluster mode requires etcd + MinIO/S3 + Pulsar/Kafka — described by practitioners as significant operational overhead. Milvus Lite runs in-process via pip for prototyping. $113M raised. Linux Foundation graduated project.
Pros
- + Handles billion-scale datasets (BigANN NeurIPS winner)
- + Full GPU acceleration (NVIDIA CUDA)
- + 12+ index types — most diverse selection available
- + Zilliz Cloud provides managed experience
- + Milvus Lite for easy prototyping
- + Linux Foundation graduated project
Cons
- - Complex self-hosted architecture (etcd + MinIO + Pulsar)
- - Heavy resource requirements
- - Schema modifications require complex migrations
- - VDBBench is Zilliz-maintained (verify independently)
Details
- Pricing Model
- open-source
- Starting Price
- $0
- Self-Hosted
- Yes
- Cloud Hosted
- Yes
- Founded
- 2017
- Repository
- GitHub →
Best For
- • Billion-scale vector search
- • GPU-accelerated similarity search
- • Maximum index type diversity (12+ options)
- • Enterprise deployments needing Zilliz Cloud
Integrations
Articles featuring Milvus
Last updated: