StarRocks Brings Lance Vector Search Into Distributed SQL
Alibaba Cloud StarRocks authors report that the Stella kernel in EMR Serverless StarRocks now integrates Lance data access and indexing into distributed SQL execution, enabling filtering, similarity search, joins, and aggregation in one query. The architecture supports global late materialization for wide result columns, and a customer case covers two billion 1024-dimensional vectors across 500 fragments. The authors claim waiting time fell to 1/40 of the former level and mixed-retrieval performance reached 16 times that of Ray + Lance, but these results are vendor-reported and not independently verified.
The integration could reduce handoffs between vector retrieval and OLAP systems, allowing multimodal AI workloads to combine similarity search with transactional joins and statistical analysis inside one distributed SQL workflow.
作者：范振，StarRocks TSC Member；阿里云开源 OLAP 负责人
在 Lance 的客户实践中，我们发现，业务对检索的要求正在向后延伸。找到相似轨迹、驾驶帧或机器人任务片段之后，还需要关联订单、模型版本和执行记录，解释这些样本为什么相似、异常集中在哪里，以及哪些数据值得进一步分析。检索命中的记录，也由此成为下一步业务计算的输入。
当检索和分析分成多个任务执行，结果集的交接、数据读取和批次等待，会拉长整个探索过程。业务希望能够直接提出问题，在同一条 SQL 中完成条件过滤、相似度检索、多表关联与统计分析，并随着问题变化继续追查。这也是阿里云 EMR Serverless StarRocks 接入 Lance 时重点解决的问题。
围绕这一需求，EMR Serverless StarRocks 的企业级云原生内核 Stella，将 Lance 的数据访问与索引能力接入分布式 SQL 执行体系，并针对混合检索、候选归并、物化策略和缓存调度进行优化。自 2025 年启动接入以来，相关能力已通过多个行业标杆客户的实际负载持续打磨。
#Alibaba Cloud#EMR Serverless StarRocks#StarRocks#Stella#Lance#Ray