Systematic Modeling for Delivery Data Governance: Building a Unified Data Foundation

The rapid growth of the digital economy creates new opportunities and challenges for enterprises, prompting a focus on data governance, breaking data silos, unlocking business value, and ensuring data security; this article shares Meituan Delivery's experience in building a unified delivery data "foundation" that connects data definition, model design, and data production.

In the introduction, the authors explain that the delivery data foundation bridges the gap between data definition and production, unifies three stages—definition, modeling, and production—to resolve data trust issues and provide a reference for practitioners.

Systematic modeling is defined as a metadata‑driven process built on dimensional modeling and pre‑governance principles; it starts with high‑level model design that decomposes business metrics into atomic and computed indicators, automatically generates detailed physical models, and then produces data processing logic, ensuring consistency from requirements to implementation.

The approach emphasizes two unifications: aligning data requirements with model design, and aligning model design with physical implementation, thereby creating a cohesive information architecture.

The need for systematic modeling arises from previous fragmented workflows where requirement management, model design, and development were disconnected, leading to ineffective data architecture management, mismatched metadata, and costly "silo" development; systematic modeling addresses these issues by standardizing the foundation.

Specifically, systematic modeling eliminates siloed development by integrating design and development tools and mechanisms, and the standardized metadata it produces eases data retrieval and comprehension for business users.

Implementation begins at the source, linking data definition, model design, and ETL development to achieve "design is development, built is obtained"; the strategy defines metrics, drives model design, constrains data processing, and creates a digital twin of the physical world, supported by tooling.

High‑level model design involves identifying business processes, technically defining metrics, and designing consistent dimensions, resulting in a unified high‑level model that guides downstream physical modeling.

Metrics are categorized as atomic (indivisible business fields), computed (derived via arithmetic on atomic metrics), derived (combining metrics with time periods and conditions), and constraints (logical tags for metric definitions).

Detailed model design translates the high‑level model into physical models aligned with data‑warehouse layers, generating DDL and DML automatically; it respects the warehouse's layered architecture (preparation, intermediate, core) and uses metadata to drive model creation.

The data‑warehouse layering includes preparation data, intermediate aggregation layers (B2, B1), and core detailed layers (B3), each supporting different analytical needs.

Metadata‑driven detailed design follows three steps: naming the physical model, automatically generating base facts from business snapshots, and deriving facts from metric constraints, ensuring alignment across definition, design, and production.

Consistency dimensions are established via a bus matrix, avoiding frequent post‑deployment adjustments and enhancing model stability and reusability.

Product implementation produces standardized DDL and DML, along with a data dictionary that maps physical tables, fields, metrics, and dimensions, facilitating downstream asset consumption.

Pre‑release checkpoints verify consistency of identical metrics from different sources, ensure business definitions match implementations, enforce compliance rules (e.g., primary key uniqueness, scan avoidance), and assess downstream impact of changes.

The summary highlights that systematic modeling improves data quality and efficiency, creates high‑quality metadata, enables a searchable data map, and supports a "service + self‑service" model that eliminates reliance on extensive SQL coding, allowing analysts to assemble reports via drag‑and‑drop.

Authors Wang Peng, Xin Xing, and Xiao Fei are members of Meituan's Delivery Data Team.