Design and Practice of Ant Group's Metric System

The speaker, Wang Gaohang, a senior technical expert at Ant Group, introduces his background in data middle platform development and outlines the agenda of the presentation.

He defines a metric from a statistical perspective, explains its role in a data warehouse, and describes the three layers of a metric system: the conceptual layer (core business concepts), the process/mechanism layer (ensuring continuity and governance), and the product layer (platforms delivering metrics). Common problems such as ambiguity, sustainability, and efficiency are highlighted.

The design section covers how to achieve concept consensus through BI‑driven or data‑architect‑driven approaches, the scope of consensus (company‑level, department‑level, team‑level), and the three semantic‑layer architectures (integrated with the data layer, independent product, or embedded in consumption tools). Domain‑Driven Design and AI‑assisted modeling are introduced to improve macro‑level abstraction and modeling efficiency.

Mechanism and workflow design aim to keep metrics fresh and maintainable. Key roles—business owner, technical owner, and consumer—are defined, and a change‑management process is presented to ensure clear responsibilities and timely notifications.

Productization discusses the indicator platform’s four core capabilities: standardized metric definition, efficient metric development, convenient consumption, and drill‑down analysis. Typical platform structures and current shortcomings (limited standardization of physical definitions, drill‑down difficulty, and consumption flexibility) are examined.

Ant’s metric platform upgrade introduces a unified term library with physical definitions, automatic calculation templates, and logical tables that provide strong lineage, improve development efficiency (10× speedup), and reduce computation cost (≈20% reduction). Materialization strategies based on ROI, frequency, and dimension usage are also described.

Practical results show nearly 30,000 derived metrics with >70% automated creation, a ten‑fold R&D efficiency gain, and significant cost reductions across Ant Group, Netbank, and Ant Security use cases. Automation rates exceed 85% in many scenarios, shortening delivery cycles from days to hours.

Future outlook focuses on leveraging large models for semantic‑layer assistance, enhancing natural‑language query and analysis, while acknowledging that full replacement is not immediate. A recruitment note for Java engineers is also included.

The Q&A section addresses physical‑caliber binding, storage formats, calculation origins, performance guarantees, and the balance between offline and online computation.