How ByteDance’s Video Processing Platform Achieves Billion‑Scale High Availability

Video Processing System Overview

ByteDance's Volcano Engine video platform supports ToB video services and products like Douyin and Xigua, handling the entire video lifecycle.

Video Lifecycle Stages

Client‑side production : Creators capture and edit videos on devices, then upload via SDK.

Cloud‑side production : Video processing and moderation run in parallel.

Cloud‑side delivery : VOD service provides playback URLs and metadata; CDN delivers the content.

Playback : Playback SDK renders video on the client.

The processing system is the core of cloud‑side production.

Challenges

Massive scale: Billions of video variants per day, heavy compute and storage demand.

Multiple business lines: Short, medium, long videos, VOD, live, RTC, education, gaming, etc.

Complex resources: CPU, GPU, FPGA, elastic resources, and hardware transcoders.

Rapid growth and peak events: Volume doubles yearly, large‑scale events stress the system.

System Goals

Meet business requirements.

Improve user experience (quality, smoothness).

Reduce cost (compute, storage, CDN).

Achieving these goals requires capabilities such as transcoding, editing, analysis, and image processing, all built on a highly available, scalable foundation.

Architecture Overview

The system is divided into three planes:

User plane : How users invoke the system.

Control plane : Interfaces for developers, operators, and support staff to manage and troubleshoot.

Data plane : Massive data for analytics, billing, and monitoring.

Four middle layers:

Service layer : Auth, task queue, template and policy management.

Workflow system : DAG‑based orchestration of media processing tasks.

Lambda : High‑availability function compute platform for resource scheduling.

BMF : ByteDance Media Framework, a dynamic multimedia processing framework.

Service Layer and Workflow

Key service components:

Service gateway : Cross‑region traffic routing, authentication, rate limiting.

Management service : Metadata management, workflow triggering, lifecycle control.

Elastic queue : Isolates business resources, configures QPS and concurrency.

The workflow engine organizes tasks as a DAG, handling dependencies, retries, and timeout detection, providing at‑least‑once execution and idempotency.

High‑Availability in Task Execution

Ensures every task eventually runs, guarantees idempotent results, and de‑duplicates repeated submissions.

Rapid Response and Recovery

Multi‑level throttling : Prioritizes critical tasks during resource shortages.

Batch re‑transcode : Quickly reprocesses affected videos after a faulty release.

System‑Level HA

Redundant middleware, downstream health checks, circuit breaking, and traffic switching strategies.

Function Compute Platform

Provides massive horizontal scaling, heterogeneous resource management, and fault‑tolerant execution for fine‑grained video processing functions.

Multi‑cluster design offers disaster recovery, automatic traffic adjustment, and seamless failover.

Control Plane – Service Governance

Each layer (gateway, scheduler, client) includes health monitoring, circuit breaking, and exception handling.

Dynamic Multimedia Framework (BMF)

BMF modularizes atomic video capabilities, supports multiple languages (C++, Python, Go), and enables dynamic registration of modules, reducing development cost and improving flexibility.

Developers can focus on individual modules, register them, and compose applications without recompiling the whole system.

Conclusion and Outlook

The architecture combines a media workflow, function compute platform, and BMF to achieve high availability, scalability, and operational efficiency. Future work aims at more intelligent, distributed scheduling where users only define processing pipelines.