Beyond Benchmarks: Real-World AI Search Performance in OpenSearch (Lucene Vs Faiss Vs JVector)
summary
Compares Lucene HNSW, Faiss and jVector inside OpenSearch using a real client index cloned into blue/green experiments, 100 queries on ~100k vectors (K=100) running on R8 2x large. Finds Lucene fastest at 1x, jVector fastest at 4x, Faiss best at 32x compression; tail latencies hover around ~0.5s and reducing K lowers latency but sharply reduces recall due to post‑filtering. Notes upgrade and plugin pitfalls (2.19→3.3 bug fixed in 3.4, nested-field bug fixed in 3.5.x) and CPU/disk, not RAM, drive compressed-vector costs.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 1:47AI search challenges in OpenSearch
- 3:12Benchmarks vs production struggles
- 8:51Debugging query latency
- 10:15JVector introduction
- 13:03OpenSearch upgrade bugs
- 15:57Nested field issues
- 19:08Lucene vs Faiss vs JVector
- 24:04Key takeaways
- 26:53Community support journey
- 27:38Q&A
related talks
55mVector Search Made Simple: Getting Started with OpenSearch for AI Applications
44mVector Search Beyond the Hype: Real-World Implementation and Performance Insights
22mComparing vector implementations in generic databases
05mKeynote: OpenSearch 3.0 and Lucene 10, Paving the Way for AI and Vector Innovation
46mAccelerating OpenSearch K-NN With FAISS, Leveraging AVX-512 for a...
43mOpenSearch and the Path to a Trillion-scale Vector Database
54mThe Debate Returns (with more vectors): Which Search Engine?
24mVector Search at Uber