Metric Management and Standardization in Didi's Data Platform

Metrics are essential measurements of business processes and outcomes, forming the core of an enterprise data warehouse. A well‑designed metric system supports fine‑grained operations and strategic decision‑making.

The presentation, delivered by Didi expert engineer Zeng Jing, covers five main topics: metric management background, Didi data product overview, metric standardization construction, and future plans.

1. Metric Management Background – The typical metric delivery workflow involves DS or operations proposing metric needs, data BP confirming metric definitions, and data warehouse engineers developing ETL. Delivered metrics are used in downstream products such as dashboards, BI reports, and analysis tools.

The basic requirements for metrics include accuracy, timeliness, consistency, and management efficiency. Didi's 1.0 data system faced scattered metric definitions, high production cost, low consumption efficiency, and inconsistent metric definitions across products.

2. Solution Development – In the 2.0 phase, Didi aims to build a standardized metric system. Two solutions were explored: a traditional metric dictionary and a metric management tool based on dimensional modeling. The tool automatically generates standardized metric names, clarifies ownership, and enforces strict entry processes.

The methodology divides business into independent blocks (e.g., ride‑hailing, two‑wheelers), then further into data domains, time cycles, business processes, atomic metrics, and modifiers. This hierarchy enables automatic generation of derived metrics and ensures consistent naming.

3. Data Product Overview – Didi’s data products include executive dashboards, business BI panels, and DS analysis tools. Two consumption models exist: traditional data‑warehouse mode (centralized metric definition, APP tables, downstream data sets) and agile BI mode (self‑service metric definition on wide tables). Both models suffer from fragmented metric management, duplicated development, and limited analytical flexibility.

4. Metric Standardization Construction – The overall goal is to decouple metrics from production and consumption, creating a headless‑BI style architecture. Three pillars are introduced: metric definition standardization, logical metric implementation, and unified metric consumption via a DSL‑based query service.

Standardized metric definitions include basic, calculated, and composite metrics, with four dimension types (dimension tables, enumerations, degenerated dimensions, derived dimensions) supporting logical snow‑flake modeling.

Process standardization automates metric entry, enforces change approvals, and implements row‑level and column‑level permissions.

Quality assurance spans ETL testing, DQC rule monitoring, and logical model validation, ensuring consistency across models and automatic cascade updates when base metrics change.

5. Unified Consumption and Future Roadmap – A unified query layer builds a metric DAG, selects optimal models based on period, engine, granularity, and efficiency, and generates federated SQL for execution. Data virtualization and MPP engines (e.g., StarRocks) provide high‑performance, low‑latency queries.

Future plans include expanding the ecosystem (experiment analysis integration, self‑service analytics, large‑model‑driven metric retrieval), continuous efficiency improvements (automation, monitoring, real‑time metrics), and deeper standardization across the organization.

The session concludes with a Q&A addressing multi‑model queries, upstream product admission policies, and the overall vision for metric standardization at Didi.