Onehouse 宣布其 Lakegres 平台新增向量搜索功能,允许直接在现有的 Apache Hudi 和 Apache Iceberg 开放表格式中通过 Postgres 端点执行向量检索。1 该方案将索引作为衍生元数据存储在表内部,无需复制数据或重写底层文件,由同一提交操作维护一致性,解决了传统方法中需将大规模数据迁移至专用数据库的问题。1
基准测试显示,在包含2.47亿个1024维向量的数据集上,Lakegres 的暖查询中位延迟(p50)为63.5毫秒,比 Databricks AI Search 快约10倍;冷查询则快约15倍。1 测试使用 AWS m8gd/r8gd 实例,每小时成本约为5.79美元(不含 Onehouse 费用)。1
Onehouse has announced a new vector search feature for its Lakegres platform, enabling direct vector retrieval on existing open table formats such as Apache Hudi and Apache Iceberg via a Postgres endpoint 1. This approach addresses the challenge of migrating large-scale data to dedicated databases by storing indices as derived metadata within the tables themselves, without requiring data duplication or rewriting of underlying files 1. The system maintains consistency through the same commit operations used for standard table updates and supports SQL similarity searches through the Postgres wire protocol 1.
In benchmark tests performed on a dataset containing 247 million vectors with 1024 dimensions, Lakegres achieved a median latency of 63.5 milliseconds for warm queries 1. This performance is approximately 10 times faster than Databricks AI Search for warm queries and about 15 times faster for cold queries 1. The benchmarks were conducted using AWS m8gd/r8gd instances, which incur an hourly cost of roughly $5.79, excluding any fees associated with Onehouse services 1.
评论
还没有评论,欢迎留下第一条。