How Y‑Tech Overcomes High‑Latency Server‑Side Video Effects with Cloud‑Native Workflows

Background

Kuaishou generates massive short‑video content daily, and special‑effects ("effects") videos are a significant portion. Y‑Tech continuously explores the intersection of computer vision, graphics, and machine learning to deliver impressive effects, which demand heavy computation for inference and rendering.

Challenges

Server‑side effects face real‑time requirements, high resource consumption, large data volumes, and bandwidth‑intensive multimedia I/O. High‑latency graphics processing conflicts with limited server resources, and the need to separate service and algorithm layers adds complexity.

Inference and rendering pipelines are complex and cannot be expressed as simple uniform APIs.

Graphics workloads involve large data and can incur seconds‑level latency, consuming substantial resources.

Multimedia I/O demands high bandwidth.

These factors require careful latency management, resource allocation, and decoupling of service and algorithm components.

Open‑Source Solutions

Two categories of frameworks are examined:

Workflow‑based Service Organization

Using message systems to decouple microservices and enable non‑blocking asynchronous calls, exemplified by Netflix Conductor. Key takeaways for Y‑Tech include independent decision services with task queues and metadata storage for tracking asynchronous tasks.

Serverless Frameworks

Knative on Kubernetes provides deployment, traffic routing, and cold‑start handling via activators. It separates request queues from business logic, offering isolation but still faces cold‑start latency for large effect containers.

Y‑Tech Server‑Side Effects Platform Design

The platform adopts a non‑blocking API with task queues, a dedicated scheduler, and process isolation for algorithms. It adds a delayed‑retry queue to reduce task loss, and uses dynamic configuration to load effect modules on demand, shrinking container sizes.

Future Plans

Further optimizations aim to improve I/O organization, reduce perceived latency, and explore richer workflow models for effect composition.

Conclusion

The Y‑Tech server‑side effects platform is operational with several services deployed, yet opportunities remain to enhance input/output handling, latency, and workflow integration, driving ongoing research and development.