Building the Digital Foundation for Intelligent Transformation with Multi-Model, Multi-Form Architecture — An Exclusive Interview with Zhang Yi, General Manager of GBASE's GBase 8c Product Division

The 17th China Database Technology Conference (DTCC 2026) was successfully held in Beijing. During the conference, GBASE hosted a dedicated GBase 8c technical forum, showcasing innovative practices spanning "from AI-native database to industry deployment." Following the event, Mr. Zhang Yi, General Manager of GBASE's GBase 8c Product Division, sat down with ITPUB for an in-depth discussion on GBase 8c's product strategy, multi-model multi-form architecture, database innovation in the AI era, and industry deployment practices.

The following is the full transcript of the interview:

Q1: What strategic role does GBase 8c play within GBASE's overall product portfolio?

Zhang Yi: GBase 8c is GBASE's next-generation native distributed database, independently developed from the ground up.

Today, AI applications are presenting unprecedented challenges to databases: large model training datasets have reached petabyte scale, straining traditional storage expansion; AI training and inference workloads fluctuate dramatically, demanding on-demand compute scaling; agent-based concurrent access drives millions-level concurrency requirements; and new scenarios like RAG and semantic search require native vector support within the database.

"Databases must evolve from being pure data storage engines into core components of AI infrastructure," said Zhang Yi. And "multi-model, multi-form" is exactly the core architectural philosophy driving GBase 8c's response to this evolution.

Q2: What progress has GBase 8c made over the past year?

Zhang Yi: Over the past year, GBase 8c has achieved critical progress in both product capabilities and market recognition.

On the technology front, we have continuously enhanced HTAP and DB4AI capabilities. In May 2026, GBase 8c V6 passed the national security and reliability assessment.

On the industry deployment front, GBase 8c has achieved large-scale adoption across finance, telecommunications, healthcare, government, and other critical industries — Luzhou Bank's credit business system achieved significant batch processing efficiency gains; a telecom operator's settlement system processes billions of call detail records monthly; and a provincial-level national health information platform center system manages 5 PB of data.

Q3: How should we understand GBase 8c's multi-model, multi-form architecture? Why did you choose this approach, and what results have you achieved?

Zhang Yi: "Multi-model, multi-form" is the core architectural philosophy behind GBase 8c's response to the challenges of the AI era.

Why multi-model multi-form? Faced with new challenges in the AI era — petabyte-scale data, elastic scaling, and millions-level concurrency — traditional database architectures can no longer address all requirements simultaneously.

By "multi-model," we mean unified support for multiple data models. GBase 8c supports unified storage for structured data, vectors, graphs, time-series, and other data types — reducing development and operational complexity and eliminating the need for users to maintain multiple separate systems for different data types.

At the storage engine level, GBase 8c offers three modes: OLTP row storage, OLAP column storage, and HTAP row-column hybrid storage.

Row storage physically stores data contiguously by row, supporting high-concurrency reads and writes with row-level locking — ideal for transaction-intensive workloads.

Column storage physically stores data contiguously by column, leveraging CU compression units and vectorized execution — ideal for complex analytical aggregation.

Row-column hybrid storage allows a single table to have both row-store and column-store physical forms simultaneously. The query optimizer automatically selects the optimal path for OLTP and OLAP workloads, with near-real-time synchronization achieved through incremental buffering and background batch row-to-column conversion.

By "multi-form," we mean flexible adaptation of deployment forms. GBase 8c offers three deployment forms:

Primary-standby — simple and efficient, suitable for scenarios with lower data volumes, where extreme single-node performance is required with data backup.

Distributed — fully redundant components with storage-compute separation, allowing horizontal scaling of compute and storage independently based on business needs — suitable for high-volume, high-concurrency scenarios with stringent data security requirements.

Storage-compute separation — built on shared object storage, supporting second-level scaling and rapid branch creation — ideal for workloads with significant fluctuations or intermittent patterns, dramatically reducing costs and better serving modern development workflows such as AI Agents.

On the multi-model, multi-form front, we have made systematic progress over the past year. Through the DBCOMPATIBILITY parameter, GBase 8c supports multiple syntax compatibility modes including Oracle, MySQL, PostgreSQL, and SQL Server. It also provides a full suite of automated migration and synchronization tools covering migration assessment, full migration, incremental sync, data verification, and reverse sync — minimizing migration and adaptation costs for users.

Q4: What has GBase 8c accomplished in terms of intelligent capabilities — both DB4AI and AI4DB? Can you share some real-world practices?

Zhang Yi: On the intelligence front, we are advancing along two parallel paths: DB4AI and AI4DB.

On the DB4AI front, the biggest innovation is native vector storage and scalar-vector fusion retrieval. Most AI applications today layer new AI capabilities on top of legacy business systems, creating demand for multi-label filtered queries combining traditional scalar data and vector data. GBase 8c achieves deep integration of vector data and relational data, supporting complex JOIN operations and vector similarity search within a single SQL statement.

Specifically, the system first generates a filter bitmap based on scalar conditions, then retrieves vector data through ANN search algorithms, returning only data that satisfies both bitmap filters and vector search criteria. If no results match, it automatically expands the candidate set and retries. Built-in support for multiple quantization indexes including HNSW, IVF, and PQ enables hybrid search combining vector similarity, scalar filtering, and full-text search.

At the storage-compute separation architecture level, the compute layer is fully stateless — nodes do not persist any data, with state streamed to the storage layer via WAL. This delivers four core capabilities:

• Second-level scaling — compute nodes can be added or removed at any time without data migration.

• Scale-to-zero — compute resources can be reduced to zero during low-load periods, with idle cost approaching zero.

• Rapid recovery — stateless nodes can be quickly rebuilt after failure with zero business interruption and second-level recovery.

• Copy-on-Write data branching based on WAL logs — creating a branch does not copy actual data; it only records a branch point in the log. The new branch shares historical data pages with the original, and only modified pages are copied on write. This data branching capability enables A/B testing, agile iteration, and rapid experimentation/rollback in AI model training — all in seconds, with fully isolated branches. The storage layer uses a distributed architecture with data sharded across multiple nodes, supporting horizontal scaling to petabyte capacity, with multi-replica mechanisms ensuring reliability.

On the AI4DB front, GBase 8c leverages the AIOps intelligent operations component, deeply integrating AI algorithms with the database kernel, and continuously exploring autonomous "self-driving" capabilities. It already provides index recommendations, intelligent SQL tuning, and automatic parameter recommendations to assist DBAs in daily operations and performance optimization. Going forward, we will continue to deepen capabilities in anomaly detection, performance prediction, and self-healing.

Q5: What truly defines an AI-native database? And what are GBase 8c's future plans?

Zhang Yi: In my view, a true AI-native database cannot be simply equated to "a traditional database with vector plugins."

From GBase 8c's practice, an AI-native database must achieve breakthroughs on at least three levels: elastic scaling and stateless computing at the architecture level; multi-model integration at the data level; and deep integration of scalar and vector queries at the query level — all while providing engineering capabilities for AI workflows.

Looking ahead, GBase 8c will continue to deepen its multi-model, multi-form architecture, with sustained investment in vector retrieval capabilities, elastic scheduling for storage-compute separation, and AI workflow support. We will further enrich native support for graph and time-series data, enhance intelligent query optimization and resource scheduling, and drive GBase 8c's evolution from a "database that supports AI" to a true "AI-native database."

Our goal is clear: to deliver stable, efficient, and cost-effective data services to users in an environment of explosive data growth and dramatic workload fluctuations.