Design and Optimization of Baidu Short Video Push System

The article introduces the short‑video Push system used by multiple Baidu apps (e.g., Haokan Video, Live, Du Xiao Shi) and explains its purpose: delivering personalized, operational, and real‑time notifications to billions of active users to boost engagement and retention.

Background : In an era of information overload, Push technology shifts from pull‑based retrieval to proactive delivery, making timely, user‑specific notifications essential for mobile applications.

Key Terminology : definitions of Message Push, Personalized Push, Operational Push, and Real‑time Push are provided to clarify the different push scenarios.

System Overview : The architecture consists of a material center, user center, personalized recall service, realtime‑API, frequency‑control (UFC) service, pre‑processing service, sending service, receipt service, and a control center (PCC). Data flows from client SDKs through user and material centers, into recommendation engines, task creation, pre‑processing, and finally delivery via cloud Push middle‑platform to vendor agents or long‑link services.

Core Modules are described in detail, highlighting their responsibilities and interactions.

Data Flow : User behavior logs are collected, transformed into tables, used by strategies to generate material and user sets, which are then recalled, packaged into tasks, sent to the Push middle‑platform, and finally delivered to devices. Receipt data is fed back for monitoring and model training.

Iterative Optimizations :

1) Personalized first/last push time estimation based on user activity windows to improve click‑through rates.

2) User‑group service redesign: replacing physical‑machine‑based user packages with bitmap‑based user groups managed via the AMIS platform, enabling logical expressions, efficient bitmap operations, and reduced memory usage.

3) Frequency‑control service overhaul: moving from static hash‑mod allocation to a dynamically scalable consistent‑hash architecture, adding whitelist handling, and supporting fine‑grained time‑slot controls.

4) Data compression using Google Protocol Buffers, which reduces payload size to about 25% of JSON and speeds up serialization, saving ~75% Redis memory for frequency‑control data.

Conclusion : Push remains a low‑cost, high‑impact channel for mobile product operations. Effective system design—covering architecture, personalization, frequency control, and efficient serialization—directly influences user experience, engagement, and operational scalability.