TencentDB TDStore Online DDL: Technological Evolution and Innovations Background & Challenges

Middle East - English

• Middle East - Arabic
• APAC - English

SHENZHEN, China, April 22, 2025 /PRNewswire/ -- Traditional single-node databases (e.g., MySQL) use OnlineDDL and third-party tools (e.g., pt-osc) to enable lock-free schema changes, but face performance bottlenecks and struggle in distributed environments. Tencent Cloud's TDStore, a financial-grade distributed database, addresses these challenges with groundbreaking innovations:

Core Technological Innovations

1.  Multi-Version Schema Mechanism

a. Introduces schema versioning to enable metadata-only modifications in seconds (e.g., adding trailing columns, extending fields). Historical data automatically fills default values, ensuring backward compatibility.

2.  Concurrency Control & State Transition

a.  Thomas Write Rule: Reduces transaction conflicts by ignoring stale writes, improving DDL-DML parallelism.
b.  Google F1 Phased State Design: Divides DDL into three stages (delete-only → write-only → final)  to ensure global consistency and smooth transitions.

3. Write Fence Mechanism

a. Validates request versions at the storage layer, allowing writes only between adjacent states to eliminate data inconsistency risks.

4.  Fast OnlineDDL Acceleration

a.  Distributed Parallel Backfilling: Splits data into SST files for multi-node parallel ingestion via bulk load, bypassing timestamp comparisons to achieve 13x performance gains (10 minutes vs. 2.3 hours).

Practices & Optimizations

1. Performance Comparison

a. Traditional Mode (single-node): 16 threads took 2.3 hours.
b.  Fast Mode (multi-node): 48 threads completed in 10 minutes, showcasing significant efficiency improvements.

2. Partitioning Best Practices

a. Large Tables: Use HASH/KEY partitioning to distribute data evenly, enabling parallel DDL execution.
b. Cold/Hot Separation: Combine RANGE+HASH secondary partitioning for rapid cleanup and elastic scaling.
c. High Concurrency: Align partition keys with frequent query fields; set partition count as multiples of node numbers.

3.  Key Parameter Configuration

a. max_parallel_ddl_degree: Increase parallel threads (≤ total node CPUs).
b. tdsql_ddl_fillback_mode:  Enable IngestBehind mode to unlock multi-node parallel acceleration.

Business Value & Future Roadmap

• Validated Use Cases: Achieved zero downtime in PB-scale financial systems, with 10x faster execution than third-party tools.
• Upcoming Enhancements:
  • Optimize partitioned table Copy Table and index backfilling for ordinary tables.
  • Support ultra-large-scale (tens of TB) workloads and hybrid HTAP architectures.

Conclusion

TDStore overcomes traditional OnlineDDL limitations through distributed architecture innovations and engineering practices, delivering high-performance, secure, and seamless schema change capabilities for financial-grade scenarios. It empowers enterprises to tackle massive data challenges effectively.

#DistributedDatabase #TencentCloud #TencentDB #TDSQL #Tencent Cloud BigData