Cassandra Vector Search Performance Tuning
summary
Claims that Cassandra 5.0 implements DiskANN via JVector and delivers competitive vector search performance, validated with a Quadrant-inspired benchmarking tool extended for Cassandra. Tests run on 3-node DS84s with 32 GB RAM and a 100M-vector dataset, varying threads from 16–100 and comparing compaction strategies, and show level compaction yields best read-heavy throughput and high precision while throughput falls as dimensionality increases; index build is faster and smaller than HNSW and supports concurrent updates.
AI-written summary · about
chapters (8)
- 0:00Introduction and motivation
- 1:34What is vector search
- 2:39Storage attached indices in Cassandra
- 3:42JVector implementation details
- 5:09Benchmarking setup and variables
- 6:13Test results and precision
- 8:01Tuning recommendations
- 9:04Future of Cassandra 5.0
related talks
35mImpact of Vector Search
40mUnder the hood of vector search with JVector
32mEnhancing Apache Cassandra Drivers with Vector Support
44mVector Search Beyond the Hype: Real-World Implementation and Performance Insights
33mEnter the Brave New World of GenAI with Vector Search
28mSponsored Session: AI-Powered Semantic Search with Cassandra: Today and Tomorrow
39mWhat's in a Vector Database?
19mHow Microsoft powers planet-scale AI apps with DiskANN