Wedge: Design and Implementation of an Advertising Experiment Configuration Platform

The article introduces Wedge, an experiment configuration platform built by Meituan-Dianping's effect advertising engine team to accelerate algorithm, engineering, and business iteration through systematic A/B testing.

Background : Advertising metrics are quantifiable, making A/B experiments essential for validating algorithm updates, large‑scale refactoring, and parameter configuration. Existing single‑key‑value configuration methods are insufficient for complex business needs, prompting the development of a dedicated platform.

Design Goals :

Support various algorithm experiment scenarios with extensibility.

Provide transparent end‑to‑end experiment configuration, usage, and result collection.

Enable flow‑based parameter configuration to reduce parsing cost.

Offer version control and quick rollback.

Deliver a concise and user‑friendly UI.

Architecture : The system is divided into four layers – Web (UI), Service (permission control, experiment management, result retrieval), Storage (data persistence), and Business (SDK integration for fetching experiment and application parameters). Figures in the original text illustrate the overall architecture.

Model Design :

1. Split‑Flow Model : Defines App, Scene, Layer, Conditional Layer, and Exp entities. Horizontal experiments occupy a single layer and are mutually exclusive, while vertical experiments span multiple layers and can coexist.

2. Experiment Hit Model : Requests are hashed first against vertical flow (using a vertical hash string) and then against horizontal flow (using a horizontal hash string). Hash priority ensures vertical experiments are evaluated before horizontal ones. Bucketing uses a default modulus of 100, guaranteeing a global 1% allocation for each bucket.

3. Application Parameter Model : Consists of global parameters and type‑specific parameters, with precedence order: experiment parameters > type‑specific parameters > global parameters.

4. Rollback Model : All configuration changes (table name, column name, type, old/new values, operation type) are recorded. Rolling back involves reversing these operations across multiple tables.

Real‑Time Effect Monitoring : Metrics such as actual CTR, predicted CTR, request PV, ad density, effective impressions, and RPS are collected via request, SPV, and CPV logs. Logs are joined on requestId and experiment path, then stored in Druid, a time‑series OLAP database, to enable efficient real‑time queries.

Results : Since its full rollout, Wedge has facilitated algorithm‑engine architecture adjustments, allowed arbitrary vertical experiments across algorithm directions, and introduced configuration audit capabilities to ensure correct releases.

Authors : Zhe Qi, Cang Kui, Liu Zheng – engineers from Meituan-Dianping's effect advertising engine team.