MILVUS-SPEC

Milvus LF AI Graduate vector DB billions vectors.

Définition

Milvus 2.x architecture cloud-native : separation Compute (Query Nodes + Index Nodes + Data Nodes) + Storage (Object Storage S3/MinIO + Message Queue Kafka/Pulsar + Metadata Etcd). Components : Proxy (handles requests), Coordinators (Root, Data, Query, Index coordinators), Worker Nodes (Query, Data, Index, Streaming Nodes). Index types : 11+ (FLAT brute force exact, IVF_FLAT, IVF_SQ8, IVF_PQ, HNSW, RHNSW_FLAT/SQ/PQ, ANNOY, DiskANN, AUTOINDEX automated selection, GPU indexes RAFT). Distance : L2, IP, Cosine, Jaccard, Hamming, Tanimoto. Hybrid search : multi-vector field search + scalar field filter integrated query DSL. Scalar fields : INT, FLOAT, DOUBLE, BOOL, VARCHAR, JSON, ARRAY. Multi-tenancy : Database concept (~Kubernetes namespace) + Collections + Partitions. SDKs : Python, Java, Go, Node.js, Ruby, .NET. Customers : eBay, Shopee, Wal-Mart, IKEA, Bilibili. Zilliz Cloud SaaS Free Tier (1M vectors) + Standard $0.30/CU (Compute Unit) hour pricing. Milvus Lite (embedded SQLite-like 2024).

Origine

Milvus open-source initial 2019 ; Zilliz fondee 2017 par Charles Xie + ex-Oracle engineers ; Milvus 2.0 cloud-native architecture 2022 ; LF AI Foundation Graduate 2023 ; $103M Series B aout 2022.

Exemple en contexte

Walmart e-commerce uses Milvus 2.x cluster ~50 Kubernetes nodes : ~1.5B product images embedded via custom CLIP model fine-tuned retail, visual similar products search ~100ms latency, IVF_PQ index 3x compression, Zilliz Cloud managed offering for production reliability.

Termes liés

• Qdrant — concurrent OSS plus light.