Hermes Push System: Architecture and Design Overview
The Hermes Push System at Xianyu separates push decisions into three coordinated services—Configuration Center for audience and material data, Task Center for timing and orchestration, and Matching Center for real‑time content ranking—leveraging MySQL, ODPS, Flink, SchedulerX, MetaQ and Alibaba’s TPP/IGraph to boost click‑through rates, double user coverage, and achieve record daily active users, while planning to add open‑page notifications and deeper AI personalization.
Background
The Hermes push system at Xianyu aims to answer three key questions for user outreach: when to send a message, which user scenario is appropriate, and what content will attract the user.
Design Idea
Hermes splits the solution into three parts:
Choosing the right sending time (When) via the Task Center.
Selecting the right audience (Who) via the Configuration Center.
Delivering the most interesting content (What) via the Matching Center.
Overall Architecture
The Configuration Center maintains core data models, providing tasks and material to the Matching and Task Centers. The Task Center determines delivery timing and orchestrates the push flow. The Matching Center selects the best material for each user based on real‑time and offline signals.
Configuration Center
Tasks, audiences, and materials are managed separately. Tasks are linked to multiple personalized materials, and each material can be targeted to specific user groups. Data is stored in MySQL and ODPS; offline sync uses ODPS, while real‑time sync leverages binlog listeners and MQ messages.
Personalized content is generated in ODPS tables and synchronized to the Matching Center via Blink (Flink‑based real‑time stream processing).
Task Center
Responsible for timing decisions, user validation, task‑material matching, and invoking the Matching Center. It uses SchedulerX for batch scheduling of personalized time tables and MetaQ for message delivery.
Real‑time triggers are driven by Omega’s behavior collection, allowing immediate push based on user actions.
Matching Center
Built on Alibaba’s TPP algorithm platform and IGraph graph database, it scores and ranks materials, performs AB experiments, and decides whether to send a message and which content to use. Fatigue control across users, tasks, and materials ensures non‑intrusive delivery.
Business Impact
After launch, click‑through rates increased by double digits, user coverage doubled, and daily active users reached a historic high. Configuration time dropped from days to real‑time, and operational efficiency improved dramatically.
Future Plans
Hermes will extend beyond push and SMS to open‑page notifications, further refine scene selection and content matching, and continue to deepen AI‑driven personalization.
Signed-in readers can open the original source through BestHub's protected redirect.
This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactand we will review it promptly.
How this landed with the community
Was this worth your time?
0 Comments
Thoughtful readers leave field notes, pushback, and hard-won operational detail here.