Tencent's SuperSQL Shatters TPC-DS Record: Unmatched Performance and Cost Efficiency

TPC‑DS: The Gold Standard for Data Analytics

The Transaction Processing Performance Council (TPC) publishes rigorous, audited benchmarks that measure a system’s ability to handle massive, complex analytical workloads. The 100 TB TPC‑DS benchmark stresses data loading, complex queries, concurrent throughput, and data maintenance, providing a comprehensive view of engineering capability.

SuperSQL’s Record‑Breaking Results

On the official TPC website, Tencent’s SuperSQL posted a composite score of 6.54 × 10⁸ , roughly ten times the prior highest record, and a unit performance cost of 11.04 yuan , about one‑sixth of earlier figures. The platform excelled in three critical phases:

Data write: 449.4 s (7.6× faster than the previous 3427.1 s).

Serial query: 177.5 s (16.5× speed‑up).

Parallel query (four tasks): 898.8 s total (11.7× faster).

Data update: 196.9 s (6.7× faster).

SuperSQL Architecture Overview

SuperSQL is Tencent’s next‑generation adaptive big‑data compute platform built on the in‑house vectorized engine TEngine . It unifies batch and streaming execution, supports native CPU + GPU heterogeneous computing, and blends SQL, Python, and multi‑modal retrieval for structured, semi‑structured, and unstructured data.

Query Optimizer : Global‑view optimization enables “optimize once, apply everywhere”.

Distributed Scheduler : Flexible support for multiple distributed paradigms and strong resource arbitration.

Universal Compute Base : A single execution layer that benefits all downstream engines.

Heterogeneous Acceleration : Native CPU‑GPU cooperation removes single‑hardware bottlenecks.

Core Optimizations that Powered the Record

1. AutoCTE – Global Reuse of Repeated Sub‑Plans

AutoCTE automatically discovers identical operator sub‑trees across different query branches, evaluates the cost of materializing versus recomputing, and extracts reusable common sub‑plans without requiring manual SQL rewrites.

2. Runtime Filter Cost Modeling

SuperSQL incorporates the construction, serialization, transmission, and probing costs of Runtime Filters into the optimizer’s cost model, allowing the planner to decide globally whether enabling a filter reduces end‑to‑end execution cost.

3. HBO & AQE – History‑Based and Adaptive Query Execution

History‑Based Optimization (HBO) leverages real‑world cardinalities and resource usage from past runs, while Adaptive Query Execution (AQE) continuously gathers runtime statistics to adjust join orders, data repartitioning, and parallelism on the fly.

4. TEngine Vectorized Execution

TEngine provides two adaptive capabilities:

Adaptive Encoding : At blocking points (e.g., Group‑by, Join, Sort), it dynamically applies Dict, Delta, RLE compression based on observed data characteristics.

Micro‑Adaptive Execution : Fine‑grained tuning of operator logic and vector batch size during execution.

5. Unified Memory Manager

A custom allocator reduces fragmentation and supports per‑core partitioning, hierarchical buffering, tree‑based quota tracking, and deterministic reclamation of memory after query completion or cancellation.

6. Adaptive Shuffle & vRDMA Integration

SuperSQL adapts shuffle partitioning to cluster scale, and its vRDMA‑based data path bypasses the kernel, provides zero‑copy transfers, and employs a proprietary congestion‑control algorithm to fully exploit bandwidth.

AI‑Driven Intelligent Diagnosis

Leveraging large‑model techniques, SuperSQL’s platform intelligence provides:

Automated performance diagnosis (20+ operator‑level issues resolved, 20.9% overall latency reduction, 46% CPU‑core‑time saving).

Workload anomaly attribution with >90% accuracy and 69.2% ticket‑resolution correctness.

Data‑acceleration and parameter‑recommendation, cutting hot‑SQL count by 25%, P50 latency by 27%, and CPU time by 66%.

DataAgent – Text‑to‑SQL Excellence

In the new Agent scenario, Tencent’s DataAgent achieved 76.23% accuracy on the Spider 2.0‑Lite benchmark, surpassing solutions from Snowflake, Oracle, and other vendors. It combines precise schema linking, mixed SQL‑Python execution, and real‑time engine verification.

Customer Value and Outlook

The benchmark demonstrates that SuperSQL’s advances in query optimization, vectorized execution, memory management, large‑scale shuffle, and AI diagnostics translate into tangible business benefits—higher throughput, lower cost, and faster insight generation for enterprise data platforms.