How AI Is Redefining Cloud‑Native Databases: The Serverless Elasticity and Evolution of TDSQL‑C

Market backdrop and challenges IDC predicts that by 2029 China's database market will reach 186 billion CNY with a 20.1% CAGR, while 66% of enterprises face AI‑driven data‑foundation reconstruction and 85% of online incidents stem from SQL issues, averaging 4.5 hours per ticket and 50% of manpower spent on tuning.

Traditional database pain points include insufficient elasticity (e.g., 10× traffic spikes during e‑commerce promotions requiring hour‑level scaling), difficult tuning for SaaS workloads with 21 k tables and 29 k joins, high storage cost, and growing operational complexity.

TDSQL‑C's breakthrough directions are twofold: Serverless architecture for automatic elastic scaling, and AI‑driven automation to free human effort for business modeling. The product positions Serverless as the foundation and AI as the engine, reconstructing elasticity, query efficiency, and management experience.

Storage‑compute separation places all compute in a resource pool while storage scales linearly with three‑replica design. Expansion of storage or compute is independent; replication switches from Binlog to physical Redolog, dramatically improving replication efficiency and achieving sub‑second cross‑region latency with true horizontal scaling and pay‑as‑you‑go billing (no‑usage‑no‑charge).

Serverless intelligent elasticity features a one‑master‑multiple‑slave architecture with a Proxy layer for dynamic traffic distribution. Two core elastic capabilities are vertical elasticity (sub‑second scaling based on dozens of metrics) and horizontal elasticity (second‑level node expansion). Billing follows a CCU model (max of CPU cores or memory) with second‑level monitoring and load‑push, enabling true consumption‑based pricing.

Predictive elasticity replaces reactive scaling (triggered at 80% CPU / 90% memory) with AI‑based forecasts. Historical load (80% of data) forms the training set, 20% the test set; a DNN model predicts scaling needs with mean‑square error < 5, achieving 80% prediction accuracy. Feature extraction includes 20 recent metrics and multi‑model alignment (LSTM, Transformer, Linear, DNN, ARIMA). The system can proactively notify of holidays or events to pre‑scale resources.

Stability guarantees include Proxy Hold to maintain connections during cross‑machine scaling (95% link‑hold rate) and kernel‑level optimizations: replacing MySQL’s double‑Mutex lock with frequent fine‑grained locks, asynchronous redo generation, and lock‑free buffer‑pool resizing, keeping query latency under 100 ms.

Releasable storage introduces trigger‑based cold‑data compression and tiered storage: hot data stays on primary storage, cold data is archived to object storage, freeing primary capacity. Write latency is in the hundred‑microsecond range, read latency for hits is also sub‑millisecond, while non‑hits are in the hundred‑millisecond range; ingestion speed averages 5 GB/s.

Real‑world case studies demonstrate the impact: a sports‑live‑streaming platform achieved 30 minutes of pre‑scale expansion, handling peak load 20% above capacity, with 30% average performance uplift and 20% resource‑consumption reduction. A securities platform using a 1‑write‑8‑read architecture maintained smooth transitions with sub‑100 ms queries and zero‑downtime scaling.

AI self‑learning optimizer addresses three traditional optimizer limits (search‑space explosion, dynamic resource changes, long‑term model training). It leverages a large‑model trained on Tencent’s internal data (WeChat, Payments) to reduce total SQL latency by 52.8% in production and 46% in TPC‑DS benchmarks, covering over 2 000 instances and 3 000+ deployed instances, handling 40 TB+ data, 21 k tables, and >10 k slow SQLs, cutting batch processing from >6 h to 3.5 h (42% reduction).

AI Navigator intelligent DB management built on Hermes Agent provides three‑layer persistent memory (short‑term context, long‑term cross‑session memory, skill library), full‑text search via FTS5, closed‑loop self‑learning (auto‑extract experience, generate reusable skills), 24‑hour daemon operation with real‑time anomaly detection and second‑level alerts, local data sovereignty via SQLite storage and five‑level permission control, model support via Tencent TokenHub, and multi‑platform messaging gateways (Enterprise WeChat, WeChat, etc.). It supports four scenarios: real‑time business insight (T+1 report to dashboard), intelligent user segmentation (30% conversion uplift), real‑time risk control (millisecond‑level defense, 50% SQL latency reduction), and AI‑enhanced optimizer.

Q&A highlights reveal that the AI optimizer works at three levels: kernel‑level plan changes (e.g., choosing Hash joins), skill‑based parameter tuning (e.g., buffer‑pool adjustments) encapsulated as Hermes Agent skills, and large‑model‑driven SQL rewriting and web‑search. The AI era shifts database usage from DBA‑centric to AI‑centric, demanding native support for multimodal data (images, sequences, vectors) and RAG capabilities without separate vector stores.

Overall, TDSQL‑C demonstrates how AI and Serverless design can reconstruct cloud‑native databases, delivering elastic scaling, zero‑downtime operations, cost‑effective consumption, and intelligent management.