DTCC 2026 | Breakthrough and Refinement: GBase 8s Self-Developed Kernel — From Adaptation to Deep Optimization

From August 20–22, 2026, the 17th China Database Technology Conference (DTCC 2026) was held in Beijing, hosted by IT168 in partnership with its two technical communities, ITPUB and ChinaUnix. GBASE (General Data Technology Co., Ltd.) showcased its full-stack GBase database innovation成果 and industry solutions at the event.

On the afternoon of August 20, the 「AI-Native · Data-Driven IT Innovation — GBase 8c Database Technology Special Session」 was held. Meanwhile, Mr. Yang Wei, General Manager of the GBase 8s Product Division, delivered a systematic presentation at the 「Database Kernel and Frontier Technology Innovation」 track, sharing the latest technical advancements in GBase 8s — including HK-Tree bitmap indexing, TAC high-availability continuity, columnar storage engine, high-concurrency optimization, and NUMA affinity.

This year's conference, themed "Integrating Data, Converging Intelligence, Leading the Future," featured 2 main venues and 16+ specialized technical tracks, with over 100 top industry experts sharing insights. Covering everything from database kernel innovation, cloud-native and distributed practices, to AI-powered data, vector databases and multi-agent collaboration, to real-time data warehousing, industry applications, and scenario optimization, the conference offered a forward-looking technical perspective and practical reference for data professionals.

An HTAP-Ready BITMAP Index and Kernel Innovations for Production Environments

Mr. Yang Wei, General Manager of GBASE's GBase 8s Product Division, delivered a keynote titled 「An HTAP-Ready BITMAP Index and Kernel Innovations for Production Environments」 at the Database Kernel and Frontier Technology Innovation track.

HK-Tree: An HTAP Bitmap Index Built for Low-Cardinality Scenarios

The presentation began with HK-Tree (High-density Key Tree), GBase 8s' self-developed indexing technology.

In the field of database indexing, B+Tree has long dominated, but it has clear shortcomings in low-cardinality scenarios: low storage efficiency and inefficient index traversal with table lookups. Traditional bitmap indexes improve storage efficiency for low-cardinality data but cannot support high-concurrency DML operations or HTAP mixed workloads — they are suitable only for cold data or offline analytics.

"Our goal was to build an index that combines high-speed writes, high-speed reads, and high-density storage all in one," said Yang Wei.

That is the design philosophy behind HK-Tree. The core approach is straightforward: optimize scattered table lookups into batch prefetching, while adding GROUP and SORTED attributes to the index, with full support for transactions and MVCC. In essence, HK-Tree can directly replace traditional B+Tree in all low-cardinality scenarios — and in wide-table, multi-column query scenarios, performance gains can be exponential.

Yang Wei presented benchmark results based on 360 million rows of data. With cardinalities ranging from 100 to 5,000, HK-Tree consistently outperformed B+Tree across inserts, updates, wide-table multi-index queries, and aggregate queries. Tests covered equality queries, range queries, multi-condition composite queries, and typical analytical operations including COUNT, DISTINCT, and GROUP BY. The data confirms that HK-Tree delivers proven performance under HTAP mixed workloads.

TAC: High Availability with Application-Transparent Failover

"Many DBAs have experienced that dreaded 2 a.m. scenario: the primary node fails, applications error out, the phone won't stop ringing, and you're scrambling to fail over," said Yang Wei. "Can we make the application layer completely unaware of the switch?"

GBase 8s' answer is TAC (Transaction Augmented Continuity).

TAC's core value lies in transparent failover at the application layer. Users simply configure any node IP in the URL — no need to distinguish primary from standby. When failover occurs, TAC ensures business continuity through three layers of protection: session continuity, transaction continuity, and current SQL connection continuity.

"Whether you're on shared-storage architecture or primary-standby, whether it's queries, DML, or DDL operations, TAC has you covered," Yang Wei added. "Even in scenarios where transactions must be rolled back, the session itself remains intact — the application never disconnects."

Columnar Storage Engine and High-Concurrency Optimization

For analytical workloads, GBase 8s introduces an in-memory columnar storage engine compatible with Oracle In-Memory Column.

"You can think of it as a high-speed snapshot of row-store data," Yang Wei explained. Columnar data serves as a near-real-time synchronized replica of row-store data, supports persistent swap in and out, and allows selective columnarization of specific columns. When analytical SQL queries arrive, the system automatically routes them to the columnar engine, delivering significantly improved query efficiency.

For high-concurrency scenarios, GBase 8s takes a systematic engineering approach — not patching individual bottlenecks, but redesigning from the architecture level.

Key measures include: breaking the blocking resource access patterns of traditional databases, decomposing and restructuring the transaction processing pipeline with extensive lock-free programming; introducing AUTO TUNE for dynamic resource configuration; and deploying rate-limiting mechanisms to prevent traffic spikes from causing cascading failures.

NUMA Affinity Optimization

In the final part of his presentation, Yang Wei addressed a frequently overlooked performance bottleneck in modern multi-socket servers — NUMA architecture.

"In multi-socket servers, memory is not a unified whole. The cost of accessing different memory regions varies depending on which CPU is doing the access. Poor scheduling can lead to astonishing performance degradation," Yang Wei explained.

GBase 8s' NUMA affinity optimization spans multiple dimensions: balancing global sharing and local access for Buffer Pool, Memory Pool, and various caches; balancing flexibility and locality in process and thread scheduling; and optimizing locality for multi-table joins and parallel operators. In one test on a 4-NUMA environment, throughput (TPmC) improved from 410,000 before optimization to 1,100,000 after — a dramatic performance gain.

Kernel self-development is a marathon. With a series of kernel-level innovations, GBase 8s addresses the stringent demands of enterprise core transaction systems. GBASE will continue to invest deeply in technology R&D, leveraging its full-stack self-developed capabilities to deliver stable, efficient, and autonomous data infrastructure for mission-critical core systems in finance, energy, transportation, and other key industries — powering enterprises' digital transformation and IT modernization initiatives.