Understanding Users Through Product: The Kernel-Level Evolution of GBase 8s

At the 2026 GBASE Technology Cloud Conference, Yang Wei, General Manager of the GBase 8s Product Division at General Data Technology, delivered a keynote speech titled "Understanding Users Through Product: The Kernel-Level Evolution of GBase 8s". He systematically presented GBase 8s' continuous breakthroughs in SQL compatibility, high availability, and kernel re-architecture as an enterprise-grade centralized transactional database. Through real-world cases such as Shenzhen Metro and a trust company, he showcased GBase 8s' large-scale deployments in critical sectors including finance and transportation.

GBase 8s is an enterprise-grade centralized transactional database independently developed by General Data Technology, supporting high-availability architectures such as shared-storage clusters and primary-standby deployments. It features four core characteristics:

• High Reliability: RPO=0, RTO<3 seconds (controlled), with TAC technology enabling transparent switchover for applications.

• Large Capacity: Maximum 16TB per table, with a single node supporting hundreds of TB.

• High Concurrency: Leveraging coroutine technology and a lock-free kernel design, it supports tens of thousands of concurrent connections while maintaining sustained performance.

• High Security: Multiple certifications including EAL4+ (Enhanced Level), Level 4 of China's Classified Protection of Information Security, commercial cryptography, and classified information system protection.

Since its first commercial release in 2009, GBase 8s has continuously deepened kernel R&D, providing long-term, stable enterprise-grade services for sectors such as finance, telecom, and government.

Kernel Re-architecture: Multiple Technical Breakthroughs

Yang Wei explained that GBase 8s has undergone a systematic kernel re-architecture, making continuous breakthroughs in high concurrency, index innovation, online DDL, consistent reads on secondary nodes, column-store engine, and NUMA affinity.

HK-Tree — An Index Designed for Low-Cardinality Data

B+Tree has low storage efficiency in low-cardinality scenarios and inefficient index traversal during table access. Traditional bitmap indexes improve low-cardinality scan efficiency but cannot handle high concurrent DML, are unsuited for HTAP, and can only be used for cold data or offline analytics. GBase 8s' innovative HK-Tree is purpose-built for low-cardinality data, featuring fast write, fast read, and high-density storage. It meets transactionality and MVCC requirements, enabling direct replacement of traditional B+Tree in all low-cardinality scenarios. In a wide-table multi-condition query with 360 million rows, HK-Tree demonstrated exponential advantages over B+Tree; aggregate queries such as DISTINCT and GROUP BY also showed significant performance improvements.

High Concurrency — From "Endurance" to "No Avalanche"

Business traffic is not constant, and sudden high-concurrency bursts can impact database service, causing not only increased response latency but also dangerous avalanches and outages. GBase 8s solves this by deconstructing transaction processing and broadly applying lock-free programming, breaking the limitations of traditional blocking resource access and improving tolerance to high concurrency. For varying loads, it introduces Auto Tune technology to dynamically adjust resource allocation, coupled with necessary throttling mechanisms to prevent avalanches. The optimization results are remarkable: not only delaying the performance degradation point under increasing concurrency, but also making the degradation curve gentle after reaching that point, eliminating avalanche-style performance collapse.

Column-Store Engine — In-Memory Column Store for Accelerated Analytics

GBase 8s is compatible with Oracle In-Memory Column syntax and its in-memory columnar capabilities, supporting persistent swapping in and out. The column store serves as a snapshot of the row store with near real-time synchronization, and analytical SQL is automatically routed to the column-store engine, allowing a single system to efficiently handle both high-concurrency transactions and real-time analytics.

NUMA Affinity Optimization

Cross-NUMA memory access within a single server can significantly degrade performance, involving locality issues in Buffer Pool, Memory Pool, various caches, and process/thread scheduling. Through systematic optimization, TPmC was improved from 410,000 to 1.1 million in a 4-NUMA environment.

Online DDL — Avoiding Business Interruption

DDL operations in production are time-consuming and often cause extended business interruptions, failing to guarantee 24×7 business continuity. GBase 8s provides broader online DDL capabilities, making DDL operations non-blocking or minimally blocking whenever possible, thus reducing operational pressure on application development and maintenance teams.

Consistent Reads on Secondary Nodes — Solving the Read-Write Splitting Dilemma

Read-write splitting is common in production, but due to replay efficiency, most databases use transaction-preference replay, which prevents secondary nodes from restoring the global consistency of the primary node—a condition second only to dirty reads. GBase 8s optimizes the core visibility mechanism to achieve consistent reads on secondary nodes while maintaining replay performance, eliminating the dilemma of read-write splitting deployments.

Strengthening Reliability: From "Usable" to "Confidently Adopted"

Yang Wei highlighted three key breakthroughs in GBase 8s reliability.

TAC Transparent Switchover
Business-Transparent Operation

He pointed out that during traditional high-availability switchovers, the new primary must complete rollforward and rollback processing. If a large transaction requires rollback, recovery can take minutes. GBase 8s' TAC technology enables the new primary to go online as soon as rollforward is complete; during rollback, reads use multi-versioning and writes wait for locks, ensuring session continuity, transaction continuity, and SQL continuity—applications are completely unaware of the switch. Under TPC-C benchmark load, switchover time is within 3 seconds.

Instant Flashback
Goodbye to Lengthy Log Rollback

Traditional flashback requires reading logs from disk and replaying the entire timeline row by row, which is inefficient. GBase 8s implements instant flashback queries, allowing immediate data retrieval at a specified time without waiting for lengthy log rollback. This is useful for both error recovery and historical data analysis.

In-Memory MVCC
Eliminating I/O Bottlenecks

Traditional MVCC relies on persistent storage, whether undo log or in-place approach, inevitably facing I/O and version-read performance issues, as well as disk bloat and I/O jitter from VACUUM. Starting from first principles, GBase 8s recognizes that version data is runtime data and does not require persistence, so it stores historical versions in memory. With agile and precise garbage collection, invalid versions are quickly cleaned, effectively solving the disk bottleneck.

Seamless Migration: Dual Compatibility with Oracle and MySQL

Regarding compatibility, Yang Wei focused on GBase 8s' in-depth investment in SQL compatibility.

Oracle compatibility now covers all common database objects—data types, heap tables, partitioned tables, temporary tables, views, materialized views, indexes, triggers, sequences, synonyms, scheduled jobs, common system views, system functions, and advanced system packages. It also fully supports PL/SQL, including conditional control statements, record types, cursors, packages, nested tables, associative arrays, varrays, autonomous transactions, exception handling, and supports advanced features such as In-Memory Column, Flashback Query, and Database Link.

For MySQL compatibility, GBase 8s supports MySQL mode isolation, the SHOW series of management commands, common session-level variables, and allows MySQL SQL scripts to run directly on GBase 8s. The compatibility scope includes data types, common functions, temporary tables, partitioned tables, indexes, stored procedures/functions, triggers, and system views.

With the self-developed MTK full migration tool and RTSync real-time synchronization tool, GBase 8s provides a complete migration service chain from resource assessment, data comparison, migration assessment, to full migration, incremental sync, and reverse sync. MTK supports diverse data sources including Oracle, MySQL, PostgreSQL, SQL Server, DB2, DM, and Kingbase, with support for resumable transfers and error recovery.

Case Practices: Validated in Transportation and Finance

Shenzhen Metro
The Largest Locally-Developed Solution in Rail Transit

The Shenzhen Metro Phase IV project adopted a full-stack, China-developed solution—GBase 8s + Kylin Linux + locally-made servers—representing the largest locally-developed deployment in the rail transit industry to date. Over 220 database instances were deployed across ACC/CLC central nodes and AFC station nodes. Daily passenger flow exceeded 10 million trips, the ACC system processed tens of millions of transactions per hour, and the CLC system ran stably under 13,000+ concurrent connections, with overall resource utilization controlled at 40%. Compared with the previous foreign product solution, construction costs dropped by approximately 60%, and ongoing O&M costs were similarly optimized.

A Trust Company
Full-Stack IT Application Innovation Practice with 125TB of Data

A trust company's inclusive finance project adopted a full-stack IT application innovation infrastructure—Kunpeng + Kylin + GBase. The system used a ShardingSphere-based sharding architecture with 5 HAC clusters and 20 database instances, totaling 125TB of data. The original system was based on MySQL, and 936,000 tables were migrated using MTK and RTSync tools. Actual benchmark TPS reached 10,000–15,000, average online transaction response time was 11ms, and timeout/failure rate was only approximately 0.02%. HAC high-availability clusters met the financial industry requirements of RPO=0 and RTO<30 seconds for business continuity.

Yang Wei's speech revealed the kernel logic behind GBase 8s' continuous evolution—from syntax compatibility to seamless migration, from TAC transparent switchover to instant flashback, from in-memory MVCC to HK-Tree indexes, each kernel upgrade addresses a fundamental question: how can a China-developed database progress from "usable" to "confidently adopted" and then to "excellent" in core transaction scenarios? The large-scale validation by Shenzhen Metro and the trust company proves GBase 8s' maturity and reliability in critical infrastructure. Looking ahead, GBase 8s will continue to deepen kernel technology to provide more stable and efficient database support for core business systems in critical sectors such as finance and transportation.