Search and Sushi; Freshness Counts
summary
Open source Vespa implements a mutable in‑memory index backed by an immutable disk index and a transaction log so updates are durable and immediately visible; attribute fields support true in-place partial updates (no full reindex) and tensors are first‑class with ONNX/LightGBM support. Measured single‑digit millisecond visibility and 50,000 updates/sec per node for simple integer partial updates, while serving at Yahoo scale (~25 billion real‑time queries/day). Background flush/merge triggers around a configurable memory target (default ~1GB).
AI-written summary · about
chapters (13)
- 0:00Introduction to Vespa
- 2:15History of Vespa
- 5:45Searching and ranking evolving data
- 7:30Real-time indexing goals
- 9:37Classic inverted index limits
- 11:43Memory evolution impact
- 12:26Mutable memory index
- 15:55Field configuration in schema
- 17:42Ranking with fresh signals
- 19:08API examples for updates
- 21:16Performance metrics
- 22:41Comparison with Elasticsearch
- 23:27Q&A session
related talks
41mScaling Search and Powering Vector-Based Retrieval With OpenSea...
43mEnhancing Query Performance Through Intelligent Data Co-location
19mEnhance Search Performance At Scale with
23mThe Faster, The Cheaper
51mIn-Memory Computing - The Big Picture
30mCaching for Speed: Practical Performance Optimization in OpenSearch & Lucene
30mIncrease text search speed & accuracy with Amazon OpenSearch Service
46mEdge Performance with In memory NoSQL