How NetEase Cloud Music Cut Costs and Boosted Performance with Data Governance

Project Background

NetEase Cloud Music serves tens of millions of daily active users and hundreds of millions of monthly active users. Growth costs are high, so fine‑grained operation and segmentation are essential. User‑profile assets help both the main service and sub‑businesses expand revenue.

Technically, the historical profile system suffered from duplicated tags across 32 tables, deep dependency chains, and multiple selection products (e.g., Muse, 诺伦, Sniper) that required re‑organization. Achieving millisecond‑level user‑group selection also demanded substantial technical refactoring.

The overall situation was strong coupling of data pipelines, high storage costs, inconsistent metric definitions, and insufficient product performance.

Project Challenges

Data challenges included massive volume, long processing chains, low timeliness, and many metric definitions. Over a thousand profile indicators required unified management to ensure high cohesion and low coupling. Task chains often spanned 7‑8 layers, reducing stability, and the daily output time (around 10 am) fell short of the 6 am requirement.

Project Solution

5‑Layer Architecture

The solution introduced a five‑layer architecture:

Data Layer : Core warehouse tables for traffic, user‑center data, content, membership, community, etc.

Logic Layer : Entity‑relationship modeling to create user‑basic, behavior, and statistical profiles, ensuring data consistency, high cohesion, low coupling, and easy extensibility (e.g., adding game entities).

Application Layer : Generates full‑profile tables and slice tables such as member, daily‑active, and monthly‑active profiles.

Product Layer : Provides tag‑factory and tag‑service capabilities for downstream products like the Magic Mirror selection tool.

Business Service Layer : Integrates with downstream products (e.g., Lingqu, Tiancheng, music‑creator operations) to enable end‑to‑end user‑group selection and push.

Tag Construction

Tags are built on demand from analysts, Magic Mirror, tag‑factory, and operations teams. By aligning tag design with warehouse layering and ER modeling, the system achieves high cohesion within entities and low coupling across entities, supporting future extensions such as game‑related tags.

Assurance System

Data‑quality monitoring covers stability, consistency, timeliness, uniqueness, completeness, and accuracy.

Task Decommissioning

A strategy‑driven, tool‑assisted process gradually phases out obsolete tasks and tables.

Magic Mirror Product

The downstream Magic Mirror product consumes the unified profile service to provide tag‑based selection for various business scenarios, linking to advertising, A/B testing, and other product placements.

Project Results

Product Value

The unified data service provides efficient tag and selection capabilities across all Cloud Music business lines, supporting user operations, online activities, A/B experiments, and ad delivery. Over 1,900 crowd‑package selections, billions of tag selections, and more than 5 million push events have been executed, covering billions of users and over 100 tags.

Governance Value

Approximately 32 tables and more than 1,000 tags were decommissioned, saving roughly ¥1.5 million in storage costs and ¥2 million in compute costs annually, for a total annual saving of over ¥3 million.

Business Value

Beyond faster push latency, sub‑businesses leveraging the unified tags acquire thousands of new users daily, translating to tens of millions of yuan in cost savings each year.