Design and Practice of Ant Group's Metric System

The session, led by Wang Gaohang from Ant Group, introduces the design and practice of Ant's metric system, highlighting his experience in data middle platform development and governance.

A metric is defined as a statistical representation of aggregate characteristics, forming the core product of a data warehouse; the metric system consists of three layers: concept, process mechanism, and productization.

Common issues include ambiguity at the concept layer, difficulty maintaining freshness in the process layer, and efficiency challenges in the product layer.

Concept consensus is achieved through BI‑driven or data‑architect‑driven approaches, with scope defined by organization, business domain, or consumption scenario, balancing consensus depth and practicality.

The semantic layer can be implemented by (1) integrating with the data layer, (2) building an independent product, or (3) embedding in consumption tools, each chosen based on company needs and governance requirements.

Mechanism and process design aim to keep metrics continuously built and refreshed, supported by productization and AI‑enhanced efficiency, and clarified by defined responsibilities for business owners, technical owners, and consumers.

Productization provides four core capabilities: standardized metric definition, rapid development, convenient consumption, and drill‑down analysis, addressing the efficiency problems of traditional platforms.

Ant's upgraded metric platform introduces a unified term library with physical definitions, automatic calculation, and logical tables that improve consistency, development speed (10×), and reduce computation cost (≈20%).

Business practice shows over 30,000 derived metrics with >70% automation, 10× development efficiency, and significant cost reductions; similar gains are reported in NetBank and Ant Security contexts, with improved security and governance.

Future outlook focuses on leveraging large models to assist semantic layer construction, enhance natural‑language query and analysis, while noting ongoing recruitment for Java engineers.

The Q&A covers physical metric binding, storage strategies, calculation origins, performance guarantees, derivation handling, and the role of offline versus online computation.