Surviving 1 Billion Vectors Without Selling a Kidney: Our Low-cost Cluster
summary
A team describes scaling a vector search project from under 2 million vectors to nearly 1 billion, facing performance, index size, and resource limits. They explain the experiments and mistakes that led them to build a custom low-cost cluster that now handles thousands of queries against over 1B vectors.
AI-written summary · about
chapters (10)
- 0:00Project and vector DB intro
- 4:44Resource issues and infra choice
- 8:41IVF and HNSW indexes
- 12:25Vector quantization for memory savings
- 14:15Switching to LanceDB disk index
- 16:05Cluster scaling and distributed search
- 20:25Solution cost and success factors
- 21:53Recommendations by load level
- 23:38Further optimization options
- 24:40Q&A
related talks
34mThe Challenges of Making Vector Search Billion-scale
35mImpact of Vector Search
26mThe Rise of Vector Databases - Lessons from the Milvus Community
44mVector Search Beyond the Hype: Real-World Implementation and Performance Insights
01h 23mVector Databases Demystified: Discover how they work
36mMilvus: Scaling Vector Data Solutions with Gen AI
35mMilvus 2.0: Building the World's Most Advanced Open-source Vector Database
04mSupercharging GenAI PostgreSQL Vector Search Optimization