How Kuaishou Automates AR Effect Quality with a Scalable Detection Service

Background

The design‑to‑release workflow for AR effects involves designers, developers, and testers, making the process long and error‑prone; manual testing of hundreds of daily submissions from external designers is infeasible, so an automated detection service was deployed to cover most potential issues.

Detection Service Architecture

The service supports internal products (Kuaishou, Yitian, overseas apps) and external creators via the Biyang Open Platform. Uploaded effect assets are sent to a task platform that creates detection jobs, which are then processed through static analysis, dynamic server‑side rendering, and real‑device testing. Results are stored in a database and forwarded to the effect producers as needed.

Task Platform

The platform receives detection requests, generates jobs, and pushes them to a Kafka queue. A dispatcher pulls jobs and assigns them to idle executors; if all executors are busy, jobs are placed in an RMQ delayed queue to be retried later, balancing latency tolerance with result reliability.

Task Execution

Static Detection analyzes effect files, checks configurations, validates file names, version numbers, redundant files, and empty frames, and estimates memory usage by inspecting images and models.

Dynamic Detection runs a DEBUG build of the app on the server to render effects, exposing OpenGL leaks, missing assets, code exceptions, and parameter misconfigurations that are often hidden by compatibility layers.

Real‑Device Testing executes the effect on a fleet of devices using a custom performance tool, collecting frame rate, memory, and CPU usage, with screenshots and recordings for reviewer inspection.

Result Notification

Because detection can be time‑consuming, results are delivered asynchronously via callbacks to the originating service (e.g., Biyang backend) or, for internal designers, through the KIM messaging tool.

Conclusion

The automated detection service now supports multiple internal teams and the open platform, catching many erroneous or performance‑problematic effects before release, reducing manual review effort and strengthening the quality gate for AR content.