BlendHouse: The Award‑Winning Cloud‑Native Vector Database Redefining Search

Recently, at the IEEE International Conference on Data Engineering (ICDE 2025)—one of the three top global academic conferences in the database field—the ByteHouse team’s paper “BlendHouse: A Cloud‑Native Vector Database System in ByteHouse” was selected for the Industry Track and received the Best Industry and Application Paper Award.

Paper title: BlendHouse: A Cloud‑Native Vector Database System in ByteHouse Authors: Zhaojie Niu, Xinhui Tian, Xindong Peng, Xing Chen Link: https://www.computer.org/csdl/proceedings-article/icde/2025/360300e332/26FZCwVQeMU

ICDE is a CCF‑recommended A‑class international academic conference, highly authoritative and influential worldwide; the inclusion of ByteHouse’s paper signifies strong recognition of its research achievements in vector retrieval.

The award‑winning paper proposes a widely applicable vector retrieval design for storage‑compute separated data‑warehouse systems. Using ByteHouse as an example, it details the complete design and implementation from storage structures and query optimization to execution, building a high‑performance vector retrieval framework called BlendHouse.

Experimental results show that this framework outperforms dedicated vector databases and existing vector‑search extensions in read/write performance.

First, a cloud‑native vector retrieval framework. BlendHouse is built on the general storage‑compute separated architecture of the relational database ByteHouse, demonstrating for the first time the feasibility of high‑performance vector search on a cloud‑native database, providing a new technical path for vector data processing in cloud environments.

Second, a universal design philosophy. The framework adopts a universal design, offering a unified vector retrieval and query chain access layer, greatly enhancing flexibility to integrate more open‑source vector indexing algorithms and better adapting to the evolving vector retrieval ecosystem.

Third, deep mixed‑query optimization. For hybrid vector and scalar queries, BlendHouse implements deep optimization strategies through a unique mixed‑query pipeline, customized optimization tactics, and vector‑semantic‑based partitioning, significantly improving execution performance for complex scenarios requiring efficient joint processing of vector and scalar queries.

ByteHouse’s vector retrieval capability is a highlight of its technology stack, crucial for semantic alignment and cross‑modal retrieval between text and images. It is already widely used in image‑text matching, product search, and multimodal large models.

For example, in a company’s “search by image” scenario handling up to 1.2 billion records, ByteHouse achieves sub‑second search speed under limited resources, demonstrating powerful vector retrieval performance.

In the AI era, the explosive growth of unstructured data such as images, audio, and video challenges traditional data processing methods. ByteHouse’s innovative vector retrieval technology provides essential support for efficient handling of massive unstructured data, becoming a core driver for enterprises’ digital transformation.