Scaling Search and Powering Vector-Based Retrieval With OpenSea...
summary
Built a highly available, multi‑tenant, multi‑AZ search platform on OpenSearch 2.3 + JDK 17 spanning 25+ clusters and ~800 nodes (~750 TB), handling ~14B writes and ~3B queries to date and peaking at ~320K writes/min and ~100K reads/min with 100–150 ms average latency. Migrated from ES6 using dual writes, Logstash reindexing and GitHub Scientist; benchmarking shows ~300% indexing improvement and ~70% lower query latency, cutting 60+ nodes and ~20% monthly infra cost. Added semantic search via a Ray GPU embedding service and Lucene HNSW (filterable) for ANN, use continuous recall 1% sampling to monitor precision, and mitigate HNSW replica nondeterminism with segment replication.
AI-written summary · about
chapters (14)
- 0:00Freshworks search platform
- 1:03Freshworks growth and use cases
- 3:09Search platform requirements
- 4:53Architecture overview
- 8:07Search service design
- 13:21Elasticsearch setup
- 16:09Elasticsearch challenges
- 17:34Benchmarking OpenSearch
- 21:25Semantic search and vectors
- 26:24Hybrid search architecture
- 29:33Vector search challenges
- 32:45Future experiments
- 34:33Scalability metrics
- 34:56Q&A
related talks
31mMaximize Resource Efficiency with Separated Index and Search Workloads
19mEnhance Search Performance At Scale with
28mRunning OpenSearch at Scale in High-Traffic Gaming Systems
32mUnleash Your Cluster's Potential with OpenSearch Benchmark
43mEnhancing Query Performance Through Intelligent Data Co-location
30mIncrease text search speed & accuracy with Amazon OpenSearch Service
19mScaling Semantic Search to Billions with Neural Sparse ANN
41mE-Commerce search at scale on Apache Lucene