Scenarios Rule! GBase’s Vector Database Development Philosophy

Recently, the vector database arena has been turbulent, with a series of major events unfolding one after another, drawing widespread industry attention. First, a technology giant announced a brand-new vector database product—a dark horse built on full-graph navigation technology—that shone brilliantly on the authoritative ANN-Benchmarks leaderboard, sweeping the top spot in six data set benchmarks. Soon after, it was reported that a well-known e-commerce platform successfully improved its recommendation system’s accuracy and user satisfaction by adopting vector database technology. These events have undoubtedly thrust vector databases into the spotlight, making them the new darling of the database field.

So, with a constant stream of vector database products and their dazzling performance claims, how can users choose the solution that best meets their business needs? Recently, Bai Junkui, Product Manager at GBase, gave an exclusive interview to ITPUB, a leading Chinese professional database and architecture community, where he revealed the “secret sauce” and “pitfall-avoidance” journey behind GBase’s vector database.

Interviewee: Bai Junkui, GBase 8a Product Manager, has 18 years of experience in the database field. He has long been on the front lines of China’s database industry, engaging in R&D, testing, technical support, L2 support management, and product management.

For professional database vendors, the technical implementation barriers of vector databases are not particularly high—capabilities such as high-dimensional support, similarity search, multiple vector fields, fast insertion and updates, multi-model joint analysis, and integration with large models are commonly required. However, applying these capabilities to real-world scenarios and maximizing their value is a continuous process of encountering and overcoming pitfalls. From engineering and productization to subsequent feature completeness, scenario adaptation, performance optimization, and stability enhancement, every step is fraught with challenges.

Drawing on two decades of database R&D experience, GBase has accumulated valuable engineering expertise and has prepared a comprehensive set of solutions to help customers address these challenges.
“We communicate deeply with our customers to understand their specific needs and scenarios, and then, based on our technical capabilities and experience, we provide tailored database solutions.”

Technical Approach and Architecture Design Considerations

Currently, there are two main technical routes for vector databases: one builds dedicated vector databases on non-relational databases, while the other adds vector database support capabilities to general-purpose relational databases. GBase chose the latter.

“We chose to add vector database support on top of general-purpose relational databases, primarily thanks to our deep R&D expertise in OLTP and OLAP relational databases, as well as our profound understanding of unstructured data management and analysis,” said Bai Junkui.

The advantage of this technical approach is that it can fully reuse existing SQL syntax, development interfaces, storage optimization, and distributed computing capabilities. Additionally, the vector database inherently possesses relational computational capabilities, enabling the fusion of vector computation and relational operations, thus providing users with more efficient and flexible solutions.

In the architecture design of the GBase vector database, a decoupled storage and compute architecture is adopted. This enables resources to scale independently—compute resources can be scaled on demand without data redistribution, significantly enhancing system availability and scalability.

“The decoupled storage and compute architecture was primarily designed to address a range of issues inherent in traditional monolithic architectures, such as single points of failure, performance penalties from data redistribution, and data redundancy,” Bai explained. Under this architecture, compute and storage resources can be scaled independently, not only increasing system flexibility but also reducing costs. At the same time, different compute resources can share the same data, minimizing redundancy and consistency risks.

Moreover, the decoupled architecture brings higher system availability and stability. By avoiding service degradation caused by single points of failure, users can enjoy more reliable data services. Meanwhile, elastic resource scaling allows users to flexibly adjust compute or storage resources according to business demands, meeting ever-growing data processing needs.

Selection Focus: Decoupled Storage-Compute and Ease of Use

When selecting a vector database, users should pay close attention to multiple key performance indicators to ensure it meets business requirements and delivers efficient, reliable performance.

First, given the massive scale of vector data, choosing a distributed, decoupled storage-compute vector database is critical. This architecture provides elastic resource scaling, ensuring the database can easily handle growing data volumes and achieve efficient resource utilization. At the same time, the decoupled design enables shared data storage, avoiding data redundancy and improving data consistency.

Second, ease of use is another key indicator. Users should prioritize vector databases that support standard SQL and development interfaces to lower development barriers and boost productivity. Such databases offer more intuitive and easily understandable operations, allowing developers to quickly master and utilize the database’s capabilities.

Moreover, supporting joint analysis of vector data and structured data is a clear business requirement. Users should choose vector databases with this capability, enabling them to process and analyze different data types simultaneously to meet diverse business scenario demands.

Finally, high security, high reliability, and easy management are factors that cannot be overlooked in the selection process. Users should opt for vector databases equipped with comprehensive security mechanisms and reliability guarantees to ensure data safety and trustworthiness. Additionally, easy-to-manage databases reduce operational complexity, improve O&M efficiency, and deliver a better user experience.

Final Thoughts

Currently, the GBase vector database has been productized and released, and we are exploring application scenarios with partners. The next step, driven by business and market demands, is to integrate with more large model providers, application vendors, and other ecosystem partners to build detailed industry-specific large model application tools and solutions, thereby empowering customers’ large model implementations.

“Intelligent operations and automated optimization will become key directions for future database development, and large model technology will play a pivotal role in this process,” Bai emphasized.

GBase will leverage large model technology to comprehensively collect and analyze database operational information, audit logs, and data access patterns. Using this information, it will adopt optimal data storage, computation, and access strategies tailored to different business scenarios, computing engines, and data content, thereby dramatically boosting database performance and efficiency.

At the same time, on the data application front, GBase will also introduce large model technology to perform feature extraction, data usage assistance, SQL review, and optimization on massive in-database data. This will help users better understand and leverage their data while improving the accuracy and efficiency of data queries and processing.

GBase is actively evolving towards a data cloud paradigm, leveraging cloud-native decoupled storage-compute technology and a unified lakehouse architecture—with unified metadata, unified compute scheduling, and unified data storage—to build a robust data cloud service platform for users. This will empower users to easily manage and utilize their data assets to meet diverse business needs.