AI Trends

Onehouse Lakegres Launches Vector Search for Hudi and Iceberg Tables

Trending on Hacker NewsUpdated Nov 19 08:42

AI summary

Onehouse Lakegres has launched vector search, enabling indexes to be stored natively in existing or newly created open Hudi and Iceberg tables and embedding vectors to be queried through a Postgres endpoint without copying the underlying data. On an approximately 1 TB table containing 247,154,006 1,024-dimensional vectors, Lakegres achieved a p50 latency of 63.5 ms for warm queries and 536.2 ms for cold queries, with Recall@10 of 0.95. Databricks AI Search recorded 655 ms, 8,242 ms, and 0.952, respectively, in the comparative test, although the two systems used different hardware and latency measurement scopes. The index is updated as data is written to the table, supports online rebuilding and optional exact reranking, and in its initial release supports filtering only through partition pruning.

Why read this

AI score 62

The original article details the storage and consistency design of vector indexes within open tables and uses tests under identical workloads to demonstrate the trade-offs among latency, recall, and build cost.

Topics

Tags

Body

Only a summary is authorized here. Continue at the validated source link.

Source

Trending on Hacker News