Milvus Review

Milvus is described as a high-performance vector database built for scale, powering AI applications by organizing and searching large volumes of unstructured data such as text, images, and multimodal content. It stores vectors alongside scalar fields (integers, strings, JSON) and supports vector search with metadata filtering or hybrid search. The README's reasons to choose it: a distributed, Kubernetes-native architecture that separates compute from storage, so query nodes and data nodes scale independently; support for all major vector index types (HNSW, IVF, FLAT brute force, SCANN, DiskANN, with quantization variants and mmap) and hardware acceleration on CPU and GPU; flexible multi-tenancy at database, collection, partition, or partition-key level, designed for hundreds to millions of tenants, with hot and cold storage; sparse vector support for full-text search with BM25 and learned sparse embeddings such as SPLADE and BGE-M3, with reranking across multiple searches; and security through mandatory authentication, TLS, and role-based access control. The project is hosted by the LF AI & Data Foundation under Apache-2.0, with Zilliz as the major contributor.

The quickest start is pip install -U pymilvus and a MilvusClient pointed at a local file, such as milvus_demo.db, using Milvus Lite (installed through pymilvus[milvus-lite]). The same client code connects to a deployed Milvus server or Zilliz Cloud by changing the URI and token. From there you create a collection with a dimension, insert data with an embedding function, and run a search. Production clusters run on Kubernetes, and the documentation covers standalone and distributed deployments, which we did not exercise.

Interaction is through the Python SDK and other client libraries, with an API of collections, inserts, and searches. Milvus Lite makes prototyping on a laptop straightforward. We did not run queries ourselves.

• Vector search with metadata filtering and hybrid search
• HNSW, IVF, FLAT, SCANN, and DiskANN indexes
• Distributed, Kubernetes-native architecture with separate compute and storage
• BM25 full-text and sparse vector search
• Multi-tenancy from database to partition-key level
• RBAC, TLS, and mandatory authentication
• Milvus Lite for local prototyping

• Scales to billions of vectors with independent read and write scaling
• Covers dense, sparse, and full-text search in one collection
• Easy on-ramp with Milvus Lite and the same client code for production
• Apache-2.0 under a neutral foundation, about 46,400 stars

• Full distributed deployment is complex and Kubernetes-oriented
• Overkill for small datasets where a PostgreSQL extension is enough
• Zilliz Cloud pricing was not readable on the page we fetched
• Another service to run and secure if you self-host

Milvus is free under Apache-2.0. Zilliz Cloud, the managed service, offers Serverless, Dedicated, and BYOC options and a free trial per the README, but the pricing page returned only its title when we fetched it, so we cannot report plan prices.

Pinecone is a fully managed, closed-source vector database service with a simple API and no infrastructure to run. Milvus is open source, can run anywhere from a laptop to a Kubernetes cluster, and offers a managed version, giving you control and portability at the price of operational complexity when self-hosted.

• AI teams building RAG and semantic search at large scale
• Organizations that want an open-source vector database with a managed option
• Multi-tenant SaaS applications

• Small projects with a few thousand vectors
• Teams without Kubernetes experience who want a self-hosted production cluster
• Those who want a general-purpose relational database