[Paper Review] Manu: A Cloud Native Vector Database Management System

Rentong Guo, Xiaofan Luan+13 more|arXiv (Cornell University)|Jun 28, 2022

Advanced Database Systems and Queries4 citations

TL;DR

Manu is a cloud-native vector database management system designed for high scalability, tunable consistency, and elastic performance in large-scale vector workloads. It decouples write and read components via a log-backed architecture using write-ahead logs and binlogs, enabling fine-grained component scaling, low coupling, and efficient multi-version concurrency control for improved evolvability and performance in real-world applications like recommendation and search.

ABSTRACT

With the development of learning-based embedding models, embedding vectors are widely used for analyzing and searching unstructured data. As vector collections exceed billion-scale, fully managed and horizontally scalable vector databases are necessary. In the past three years, through interaction with our 1200+ industry users, we have sketched a vision for the features that next-generation vector databases should have, which include long-term evolvability, tunable consistency, good elasticity, and high performance. We present Manu, a cloud native vector database that implements these features. It is difficult to integrate all these features if we follow traditional DBMS design rules. As most vector data applications do not require complex data models and strong data consistency, our design philosophy is to relax the data model and consistency constraints in exchange for the aforementioned features. Specifically, Manu firstly exposes the write-ahead log (WAL) and binlog as backbone services. Secondly, write components are designed as log publishers while all read-only analytic and search components are designed as independent subscribers to the log services. Finally, we utilize multi-version concurrency control (MVCC) and a delta consistency model to simplify the communication and cooperation among the system components. These designs achieve a low coupling among the system components, which is essential for elasticity and evolution. We also extensively optimize Manu for performance and usability with hardware-aware implementations and support for complex search semantics.

Motivation & Objective

• To address the limitations of traditional relational DBMS designs in handling modern vector database workloads with high-scale, unstructured data.
• To enable tunable consistency and performance trade-offs for diverse application requirements, moving beyond strong or eventual consistency alone.
• To achieve fine-grained elasticity and component-level scalability by decoupling functional and hardware layers in vector database systems.
• To optimize for performance and usability through hardware-aware implementations and support for complex search semantics.
• To support long-term evolvability and failure isolation in cloud-native environments through a modular, log-centric system design.

Proposed method

• Exposing write-ahead logs (WAL) and binlogs as backbone services to provide time semantics and decouple components.
• Designing write components as log publishers and read-only analytic/search components as independent subscribers to the log stream.
• Employing multi-version concurrency control (MVCC) and a delta consistency model to simplify inter-component coordination.
• Decoupling vector search, index building, and log archiving as separate functional components for independent scaling.
• Utilizing hardware-aware optimizations for performance, including SIMD, cache efficiency, and support for GPUs/FPGAs.
• Implementing a modular, extensible architecture that allows co-processors (e.g., for primary key or text search) to subscribe to the log stream.

Experimental results

Research questions

• RQ1 How can a vector database achieve high elasticity and evolvability while supporting tunable consistency in cloud-native environments?
• RQ2 What architectural patterns enable low coupling between write and read components in vector databases without sacrificing performance?
• RQ3 How can a log-based backbone simplify inter-component communication and enable flexible consistency models?
• RQ4 What performance optimizations are most effective for large-scale vector search and index management in distributed systems?
• RQ5 How can vector database systems support complex search semantics and heterogeneous workloads through component decoupling?

Key findings

• Manu achieves high scalability and elasticity by decoupling functional components such as write, search, and index building, enabling independent scaling and evolution.
• The log-backed architecture with MVCC and delta consistency enables tunable consistency, allowing users to configure visibility latency for newly inserted data.
• Hardware-aware optimizations, including SIMD and cache-aware designs, significantly improve performance for vector search and index operations.
• The system supports complex search workloads, including multi-vector and hybrid (vector + attribute) search, through extensible co-processor integration.
• Manu has been successfully deployed in production for diverse applications such as recommendation, multimedia, language processing, medicine, and security.
• The system demonstrates strong performance and scalability across three real-world application scenarios, validating its design goals.