Design and Practice of 360ShuKe Risk Control System Architecture

In this talk, Zhu Jie, Vice President of Risk Technology at 360ShuKe, presents the design and implementation of the company's risk control system, covering its overall architecture, credit data lifecycle management, and indicator computation.

The architecture consists of three layers—user, product integration, and risk—communicating via RPC, with service layers for process, calculation, and execution, and utilizes multiple databases such as MySQL, MongoDB, and HBase to handle massive data volumes.

The credit approval workflow includes pre‑screening, initial review, final review, and decision stages, supporting both real‑time (sub‑500 ms) and asynchronous processing using responsibility‑chain patterns.

Data ingestion draws from internal, external, and behavioral sources, transformed into features through an automated pipeline; the platform employs Kafka, Flink, Spark, and Kudu for reliable, high‑availability processing and monitoring.

Real‑time indicator services provide fast, accurate variable calculations for downstream models and decisions, while the feature platform evolves through two major upgrades to improve scalability, reuse, and performance.

Challenges such as third‑party data loss, rapid model iteration, and feature drift are addressed by a shadow system, feature monitoring platform, and experimentation environment that enable risk pre‑detection, faster UAT, and efficient model re‑training.

The presentation concludes with a Q&A and thanks to the audience.