Vector Search at Uber
summary
Uber deploys Apache Lucene 9 HNSW as the vector engine and builds a three-layer index—base, snapshot, live—to support live vector updates with lock-free in-memory structures, tombstones and an off‑heap skip list, ingesting from Kafka/Flink for true real-time freshness. Geo sharding (H3), static ranking and early termination scale the system for Eats and Maps, where geo-semantic search measurably raises relevance. External GPU work reports 10–50x speedups (NVIDIA) and Lucene CUVES reports ~2–4x latency gains.
AI-written summary · about
chapters (9)
- 0:00Introduction to vector search
- 1:13Uber search use cases
- 4:01Real-time ingestion challenges
- 6:08Platform architecture overview
- 9:44Vector search benefits at Uber
- 11:30HNSW algorithm in Lucene
- 12:38Live updates in vector search
- 20:08GPU acceleration for vector search
- 22:16Future work and Q&A
related talks
05mKeynote: OpenSearch 3.0 and Lucene 10, Paving the Way for AI and Vector Innovation
18mAdvancing Search at Uber with OpenSearch
15mAdvancing Search and Observability at Uber with OpenSearch
01h 09mVector Search: Ask Me Anything!
27mUnified AI & Relational Power: Scalable Distributed Vector Search with
44mVector Search Beyond the Hype: Real-World Implementation and Performance Insights
54mThe Debate Returns (with more vectors): Which Search Engine?
43mOpenSearch and the Path to a Trillion-scale Vector Database