Design and Challenges of Kuaishou Y‑tech Server‑Side Effects Platform

Kuaishou, one of China's leading short‑video communities, generates massive video content daily, with special‑effects ("特效") videos playing a significant role. Y‑tech continuously explores the intersection of computer vision, graphics, and machine learning to deliver impressive effects, but these effects demand heavy computation and large resource consumption.

Because some effects are computationally intensive and end‑device resources are limited, server‑side processing is required. The main challenges include real‑time requirements versus high latency of graphics‑heavy inference and rendering, large data volumes, bandwidth demands, and the need to separate service and algorithm layers to reduce distributed‑system complexity.

To address these issues, the team investigated open‑source workflow orchestration frameworks such as Netflix Conductor, which provides non‑blocking asynchronous task scheduling and metadata management, and serverless platforms like Knative on Kubernetes, which offer request queuing, traffic control, and cold‑start handling.

Building on these concepts, Y‑tech designed a backend architecture that introduces a task queue with delayed‑retry capabilities, leverages dynamic configuration for process isolation, and isolates effect algorithms from the service layer to improve stability and resource efficiency.

The platform now supports non‑blocking APIs, task scheduling, and state tracking, but still faces difficulties such as limited resources for high‑latency effects, retry‑induced latency, and large container images causing slow startup.

Future work includes further optimizing input/output handling, reducing user‑perceived latency, and exploring richer workflow models for effect composition. The Y‑tech team continues to recruit talent in computer vision, graphics, multimodal AI, and platform development.