OdinANN: Direct Insert for Consistently Stable Performance in Billion-Scale Graph-Based..
summary
OdinANN implements direct insert into an on-disk graph index (2x SSD overprovisioning) and claims billion-scale stability while ingesting up to 100 million vectors per day; it reduces memory (buffered deltas needed 100 GB for 3% inserts) and cuts write amplification by grouping neighbor updates into page writes via update combining. Concurrency uses approximate concurrency control (record-level, short critical sections). Evaluation shows stable search latency and throughput at ~1,000 inserts/sec and >5,000 searches/sec with preserved accuracy.
AI-written summary · about
chapters (8)
- 0:00OdinANN introduction
- 1:24Large-scale vector insert challenges
- 4:21Direct insert benefits
- 5:23Update combining for SSD writes
- 7:30Concurrency control mechanism
- 9:40OdinANN system overview
- 10:01Evaluation and performance
- 11:05Conclusion and future
related talks
45mBehind the hype: managing billion-scale embeddings in Elasticsearch and OpenSearch
28mAerospike's Journey Through Innovations, Sustainability, and Roadmap
19mHow Microsoft powers planet-scale AI apps with DiskANN
28mThe Memory Efficient Indexing for Vector Databases
38mLight and Adaptive Indexing for Immutable Databases
26mOpenSearch and the Path to a Trillion-s...
48mThe power of AI and Copilot for Azure Databases
03h 08mAzure Cosmos DB Conf 2025