Summary

What it is

A distributed vector database built for billion-scale similarity search, using a microservices architecture with SSD caching for hot data and native S3 cold storage offload.

Where it fits

Milvus is the enterprise-scale vector database for organizations that need to search billions of vectors. Its S3 integration for cold data offload and log-based write-ahead design make it the choice when scale exceeds what single-node vector databases (Qdrant, Weaviate) can handle — at the cost of significantly higher operational complexity.

Misconceptions / Traps

Milvus is a distributed system with significant operational complexity. Running it requires etcd, MinIO or S3, and Pulsar or Kafka — not a single-binary deployment.
S3 is used for persistent storage and log backup, not as a live query tier. Query performance depends on in-memory and SSD-cached segments, not S3 latency.
The microservices architecture enables scaling but introduces failure modes absent in simpler vector databases. Expect to invest in monitoring and operations.

Key Connections

depends_on S3 — uses S3 for persistent object storage of segments and logs
scoped_to Vector Indexing on Object Storage — billion-scale vector search over S3 data
solves Cold Scan Latency — hot vector caching with durable S3 persistence

Definition

What it is

A distributed vector database designed for billion-scale workloads. Uses a microservices architecture with SSD-backed caching for hot data and native offloading of cold vectors to S3-compatible object storage.

Why it exists

Single-node vector databases hit memory and throughput ceilings at billion-vector scale. Milvus distributes the index across a cluster, uses tiered storage to keep hot vectors in memory while parking cold vectors on S3, and scales compute independently from storage.

Primary use cases

Billion-scale embedding search for enterprise RAG systems, tiered vector storage with S3 cold tier, distributed similarity search across massive embedding collections.

Recent developments

Latest signals

Zilliz productized the lakehouse turn: Vector Lakebase (June 21) and the Loon storage engine (June 29, 2026). Vector Lakebase decouples compute into three modes — tiered real-time serving (in-memory 1,000+ QPS single-digit-ms down to memory+NVMe+S3 at 10–50 QPS), minute-billed on-demand search (vendor benchmark: 1B 768-dim vectors at 10 hrs compute/month ≈ $318 vs ~$4,937 on comparable serverless), and batch analytics. Underneath, Loon is a lake-native storage engine built on the Vortex columnar format — vendor-reported point reads pull ~0.07 MB vs 9.4 MB for Parquet (a ~135× read-amplification reduction), attacking exactly the Parquet-row-group problem that kept vectors out of the lake. Per Zilliz — Vector Lakebase launch and Milvus — Why we built Loon.
Milvus 3.0-beta: External Collections — zero-copy queries over the data lake (May 9, 2026). The 3.0-beta line extends Milvus beyond its own storage: External Collections query Parquet/Iceberg files on object storage in place — no ingestion copy — plus Spark snapshot reads, richer retrieval, and a more expressive schema. External Collections accept parquet, vortex, lance-table, and iceberg-table sources via bucket URIs. The beta also ships MVCC-style Snapshots (point-in-time read-only views over existing segments, no copy — batch jobs alongside live ingest), kernel-level multi-field ORDER BY + server-side GROUP BY aggregations, and NULL-vector support with online AddField. Full 3.0 GA is targeted for end of 2026. Per Milvus release notes and Zilliz release notes.
Architecture note: Woodpecker WAL replaced Kafka/Pulsar (2.6 line). Milvus 2.6 swapped its message-queue dependency for Woodpecker, a WAL implemented directly on object storage — one less distributed system to operate, and the write path now shares the S3 substrate with the data. Per Pinecone vs Weaviate vs Milvus vs Qdrant 2026 (dev.to).
GPU indexing: NVIDIA CAGRA builds a 10M-vector index in ~45s vs 18–22min CPU HNSW. Milvus remains the only major open-source vector DB with native CAGRA support; GPU-flat search returns ~2ms on 10M vectors vs ~150ms CPU. Per Self-Host Vector Databases on GPU Cloud (Spheron).
Latest release: v2.6.18 (current as of June 2026). GA on the 2.6 stable line (the 3.0.0 line remains beta); adds nullable vector support and element-level search on Struct fields. Tracking the upstream stable release line. Per milvus-io/milvus releases.
Milvus v2.6.15 (April 24, 2026) — focus on stability + recovery time. Per the milvus-io/milvus releases page, v2.6.15 ships faster MixCoord recovery, optimized search and query filter performance, and 20+ bug fixes. Go runtime upgraded to 1.25.8 for CVE-driven base-image refresh. The release pattern — frequent point releases focused on operational shapes — signals Milvus is now firmly in "production-hardening" mode rather than "feature land-grab."
Benchmark positioning: 6ms p50 latency on 1M-vector workloads. Per the SaltTechno vector DB benchmark (February 2026) — 1M vectors, 1536 dim — Milvus posts 6ms p50 / 35ms p99 latency with 8 index algorithms (including GPU). The HolySheep showdown vs Qdrant + Weaviate (April 2026) measures Milvus 2.4 at cold p50 28ms / p99 145ms, warm p50 9ms / p99 31ms, with bulk-insert of 1M vectors in 4m 12s and a 99.82% success rate under sustained 1000 QPS. Source mix note: these are third-party benchmarks rather than official Milvus measurements; treat them as positioning data, not load-bearing performance contracts.
v2.6.16 shipped GA May 13, 2026 with a compaction fix targeting delete-heavy workloads, filling the point-release gap between v2.6.15 (April 24, faster MixCoord recovery) and the current v2.6.18. Per Vector Database News May 2026 (RankSquire).
Zilliz Cloud's GA of the Milvus 2.6.x line (announced January 2026) added JSON Shredding + JSON Path indexing (vendor claim: up to 100x faster metadata filtering), hybrid BM25 full-text search (vendor claim: up to 7x faster than Elasticsearch on selected datasets), plus spatial data types, timezone-aware timestamps, INT8 vectors, and nested structures. Vendor announcement; performance multiples are Zilliz's own figures, not independently benchmarked. Per Zilliz Milvus 2.6.x GA announcement (PR Newswire).
Beyond 3.0 GA (targeted end of 2026), the published roadmap adds unified Tensor/StructList data types, geo data support, Spark/DuckDB/DataFusion query integration, "More Like This" (MLT) search, BLOB/TEXT field types, and active-standby HA — on top of the already-covered dual Parquet/Vortex storage and Woodpecker WAL. Per Milvus Roadmap (official docs). Sources: Milvus release notes — 3.0-beta · milvus-io/milvus releases (GitHub) · Vector Database News May 2026 (RankSquire) · Vector Database Performance Benchmark 2026 (SaltTechno) · Vector Database Showdown: Milvus vs Qdrant vs Weaviate (HolySheep, April 2026)