Building Tencent Xinge: Architecture and Practices for Massive Mobile Push Service

Author: Gan Hengtong, who joined Tencent TEG Data Platform in 2011, focuses on large‑scale data platforms and push service backend development. He is currently building the Xinge precise push system, covering intelligent grouping, real‑time push, and effect tracking.

The talk shares the construction of Tencent Xinge, a massive mobile push service. Push is a critical channel for reaching mobile users, but achieving hundred‑billion‑level message delivery and post‑push effect tracking poses significant technical challenges. Fine‑grained, interest‑based operations also require deep data and machine‑learning capabilities.

Push System Construction

The system consists of three main blocks: terminal, backend, and cloud‑native governance. Value‑added services include precise‑push workflow, data platforms, foundational support, and visual operation tools. The push process involves three steps: retrieving target users, selecting appropriate channels, and delivering messages to terminals.

Terminal

Key challenges on the terminal side are service keep‑alive and push delivery rate. Xinge integrates manufacturer channels (e.g., Xiaomi, Huawei, Meizu) and uses a shared‑channel approach to reduce power and traffic consumption while improving service survival. Cloud‑controlled channels enable seamless configuration updates and channel switching. Accurate statistics collection is also critical, requiring awareness of permission status and actual display on diverse devices.

Crash reporting, alarm, and online repair platforms ensure terminal quality.

Backend

The backend architecture includes terminal, access layer, logic layer, storage, data‑analysis platform, and message gateway. Real‑time effect tracking and multi‑dimensional analysis are provided by a data‑analysis platform built on Docker‑based cloud infrastructure, delivering high performance, scalability, and operational ease.

A typical use case: an e‑commerce app pushes a promotion to male users in Guangdong at 8 pm. The system supports various push types (single, batch, tag, group) and routes messages through manufacturer and custom channels, followed by data reporting.

Performance improvements focus on faster message delivery and operational efficiency. A partitioned bitmap system enables real‑time distributed user group selection and channel routing, completing billion‑level user filtering in hundreds of milliseconds.

Message dispatch relies on a centralized CKV (NoSQL/Redis‑like) store for IP/port information. The system separates upstream registration (ensuring high success rate) from downstream delivery (handling million‑plus QPS during peak pushes) using consistent hashing and cache managers.

Cloud‑Native Governance

Operations have shifted to Docker‑based CI/CD and a DevOps workflow with MySQL as a static/dynamic configuration center, reducing release time from hours to minutes. The cloud‑native approach ensures version consistency, environment isolation, and resource virtualization.

Value‑added services include precise push (rule engine, filtering, estimation models, deep learning, online learning, transfer learning), multi‑dimensional real‑time analytics, and a visual machine‑learning platform that simplifies model deployment for recommendation, OCR, NLP, etc.

The platform supports BRNN algorithms and provides a drag‑and‑drop interface for building models, reducing development cost.

Overall, Xinge integrates message push, data analysis, data operation, and business intelligence into a single service, supporting use cases such as improving game retention through precise targeting and attribution analysis.

Q&A

Q: How does Xinge handle complex filtering rules for target users?

A: Basic filters are mapped to tags; complex logic becomes combinations of tag “AND/NOT” operations, executed with sub‑millisecond latency even on billions of users.

Q: Is there performance difference among different tags?

A: No, tag processing is uniformly sub‑millisecond, independent of tag type.

Q: Are there plans to merge Xinge’s channels with WeChat’s channels?

A: No, push and IM are separate business domains; their underlying channels remain independent.

-END-