Infinigraph: Vector Data at Scale
summary
Neo4j ships InfiniGraph: a sharded property database that moves node and relationship properties to separate property shards, keeping a lightweight graph shard holding IDs and types. Using LDBC SF1000 with sentence-transformers and the Google embedding demo to produce 220 million vectors of 256 dims, an 8-shard setup stores each property shard at ~78 GB while the non-sharded vector dump is 774 GB. Adding a vector index (recommended >10k vectors) lets the vector index parallel search run across 8 property shards and yield comparable or better latency than a single large machine; an unexpected nearest neighbor example returns “Sigmund Freud.”
AI-written summary · about
chapters (6)
- 0:00Introduction to talk
- 1:06Why shard your database
- 3:13InfiniGraph architecture explained
- 9:15Performance with vector data
- 17:44Vector search query demo
- 21:58Summary and conclusion
related talks
39mWhat's in a Vector Database?
35mImpact of Vector Search
01h 23mVector Databases Demystified: Discover how they work
48mVectors are the new JSON
27mUnified AI & Relational Power: Scalable Distributed Vector Search with
32mDistributed Vector Databases - What, Why, and How
52mpgvector Newbie to Expert
51mVector data in Postgres: Size, TOAST, Filters and Performance