Kunpeng Advertising System Architecture and Operational Design for Xianyu

Background Xianyu’s rapid growth created high‑traffic scenarios such as homepage feeds that multiple business lines compete for. Legacy development and operation models could not guarantee PV targets, conversion efficiency, or fast rollout of diverse operational tactics, leading to high trial‑and‑error costs.

Idea The goal is to maximize global traffic conversion under a fixed total traffic volume by globally regulating flow. The solution focuses on: ensuring multi‑party delivery goals in a single scene, optimizing cross‑scene conversion, and enabling low‑cost, rapid business iteration through engineering empowerment.

Kunpeng System Construction A three‑level abstraction (activity → scene → material) was introduced. Scenes (e.g., search result feeds, pop‑layer, query intervention) define the placement stage; each scene can host multiple templates for material management. Operators configure reusable capabilities (UV fatigue filter, keyword filter, platform filter, version filter) via the UI, create activities, assign audiences and schedules, and launch material delivery.

DataFetcher Extension Subsystem Business lines can subclass the DataFetcher base to fetch real‑time data from remote services, convert it to domain objects, and let the underlying platform handle RPC concurrency, resource isolation, and monitoring. This enables parallel development and eliminates the previous bottleneck where only scene owners could implement delivery logic.

Filter Subsystem A modular filter library provides pre‑deployment constraints such as platform, version, gray‑scale ratio, strict/fuzzy keyword matching, page number, and UV fatigue. Custom filters can be created by extending the MatFilter base, allowing cross‑business reuse.

Dujiangyan Algorithm Platform In collaboration with the DAMO Academy, a generic mixing algorithm was built to allocate traffic globally. Offline data pipelines feed back exposure and click metrics (including unique business IDs) to continuously improve the model. Deployment of the Kunpeng system across multiple Xianyu scenes has increased PV‑CTR by 60%–100% and reduced UV‑cost by ~40%.

Group Management & Approval Flow Each user belongs to a business group; group members can only see their own materials. Any change triggers an approval workflow, preventing accidental cross‑group modifications.

Roles After Integration Scene owners handle Kunpeng integration, business developers implement extensions on DataFetcher, and operators configure materials and select base components in the console, enabling self‑service adjustments without developer involvement.

Results The system has been rolled out to homepage feeds, gold‑position slots, search results, recommendation lists, etc., achieving 60%–100% conversion uplift and >50% reduction in time‑to‑market for new business requirements.

Outlook Future work includes expanding vertical algorithm capabilities, more flexible extension points, and broader cross‑scene traffic optimization.