What database should I use for millions of vectors with low latency?

For millions of vectors with low latency, the best choice is usually a dedicated vector database rather than a general-purpose database with a vector extension.

Good options

• Pinecone – very popular for low-latency vector search at scale, managed, easy to operate.
• Weaviate – solid open-source/managed option, good filtering + hybrid search.
• Milvus – strong for large-scale vector workloads, open-source, flexible, more ops-heavy.
• Qdrant – lightweight, fast, great filtering, simpler to run than Milvus.
• FAISS – extremely fast library, but not a full database; best if you’re building your own service.

Quick recommendation

• If you want easiest production setup: Pinecone
• If you want open source + good performance: Qdrant or Weaviate
• If you need very large scale and control: Milvus
• If you’re embedding inside an app and can manage infra yourself: FAISS

What to look for

For millions of vectors and low latency, make sure the DB supports:

• Approximate nearest neighbor (ANN) search
• Filtering by metadata
• Index types like HNSW, IVF, or PQ
• Fast upserts
• Horizontal scaling / sharding
• Persistence and backups

Important note

If “millions” means maybe 1–50 million vectors, Qdrant, Weaviate, Pinecone, or Milvus are all reasonable.
If you need sub-10 ms latency at high QPS, the implementation and index tuning matter as much as the database choice.

If you want, I can recommend one based on your:

• vector dimension,
• QPS,
• update frequency,
• cloud/on-prem preference,
• and whether you need metadata filtering.