Building intelligent analytics at scale
summary
e6data’s SQL engine runs unified SQL plus vector search by doing pre-filtering (SQL first) so a 500M‑row reviews table drops to 96M by category then to 19M by rating, yielding ~6.8GB of binary quantized embeddings 32× smaller that fit a node; global quantized search over‑queries ~2× to get 5,420 candidates and then rescoring on full vectors produces the true top results, a representative query takes ~8s vs 5.5s without vectors; embeddings live in a blitz lakehouse format on S3, search runs on CPU and currently omits heavy vector indices.
AI-written summary · about
chapters (10)
- 0:00E6 Data and lakehouse analytics
- 6:25Unify SQL and vector search
- 11:44Prefiltering vs postfiltering
- 16:22Memory and compute challenges
- 23:11Binary quantization and ranking
- 28:04Storage I/O for embeddings
- 30:10Vector indices and production
- 33:45Performance on 500M records
- 35:09Future roadmap
- 36:16Q&A
related talks
27mUnified AI & Relational Power: Scalable Distributed Vector Search with
33mBringing Vector Search to OLTP
45mBehind the hype: managing billion-scale embeddings in Elasticsearch and OpenSearch
32mRetrieval Augmentation and Semantic Search at Scale
27mMachine Learning for Advanced AI Search
15mUnlock the Power of Your Data for Intelligent AI Applications
34mThe Challenges of Making Vector Search Billion-scale
30mVector Search Made Simple: Getting Started with OpenSearch for AI Applications