NrtSearch: Yelp’s fast, scalable, and cost-effective open source search engine
summary
Yelp builds NrtSearch a Lucene-based engine that uses Lucene segment replication and a primary‑only writer that uploads segments to S3 while replicas pull bytes and serve queries via gRPC. Deployed as stateless Kubernetes deployment microservices, nodes bootstrap in seconds, remove replica merge CPU work, and yield up to 50% cost reduction and ~30–50% latency improvements p50–p99. Production issues include gRPC fan‑out, segment transfer tuning, and fixes for concurrent segment slicing called virtual sharding, and plugins add vector and neural features.
AI-written summary · about
chapters (6)
- 0:00Introduction and motivation
- 3:28Elasticsearch challenges at Yelp
- 8:02Designing NrtSearch architecture
- 12:14Performance gains in production
- 17:51Future plans and open source
- 19:56Q&A session
related talks
01h 48mThe Debate: Which Search Engine?
36mWhat’s New in the OpenSearch Project and Ecosystem
12mKeynote: OpenSearch, Innovation and Community
45mOpenSearchCon Keynote Speakers
40mInnovating in Open Source in Your Enterprise
01h 03mThe OpenSearch Community, Building the Future Together
42mContribute to OpenSearch: Build Your Skills While Advancing Open Technolo...
13mOpenSearch Project Roadmap 2022 and Beyond