Exploring Parallel Query in TDSQL-C: Architecture, Performance, and Future Directions

The talk introduces TDSQL‑C, a cloud‑native relational database built on physical log replication and shared storage, emphasizing its high compatibility with MySQL and ease of migration.

It identifies the challenge of large‑scale analytical queries that cannot fully utilize idle CPU cores due to MySQL's single‑threaded SQL processing, and proposes query‑level concurrency (parallel query) to split and distribute computation across multiple cores.

Performance demonstrations using TPC‑H benchmarks show varying speedups across queries, explained by Amdahl's Law, which relates overall acceleration to the parallelizable portion of the workload.

The technical foundation is described as task decomposition and data partitioning, where operators such as hash joins and aggregations are split into parallel subtasks, with coordination handled by a scheduling framework that constructs a task graph from the original execution plan.

Parallel execution relies on a coordination thread to dispatch tasks, while worker threads execute operators and report status; the framework integrates with MySQL's exception handling to ensure robustness.

Data partitioning leverages InnoDB's B+‑tree structure, recursively dividing index ranges to balance workload among threads, and data exchange between threads is performed via bridge operators forming a data network.

Future work includes expanding parallel query coverage, deeper performance optimizations, mixed row‑column scheduling, and improving SQL execution transparency for easier troubleshooting.