Streamline: custom video pipelines with Cloudflare Stream and Workers
Cloudflare launched Streamline, a developer playground demonstrating custom video pipelines using Workers, Containers, and media protocols to modify livestreams or hosted videos with features like burned-in subtitles.
Cloudflare Stream already supports standard broadcasting, but Streamline introduces a developer playground to build custom video pipelines for tasks such as adding dynamic annotations or burned-in subtitles to livestreams or hosted videos. The system leverages Workers for control signaling, Containers for long-running media processing, and media protocols for input and output, enabling real-time video manipulation and immediate publishing. Developers can deploy Streamline to Cloudflare’s Developer Platform and test it on their own projects, with the option to run it locally using Docker during development.
The Streamline architecture consists of two main components: a Media Engine hosted in a Container for processing and input/output handling, and a controlling Application built with Workers for session management. The Media Engine can ingest RTMPS streams, Cloudflare Stream HLS manifests, or direct video sources like webcams, while the Application can start, configure, observe, and stop media sessions. The system ensures continuous processing even if the controlling Worker disconnects, using Durable Objects for orchestration and overriding container activity expiration to maintain long-running pipelines.
A Streamline deployment allows the controlling Worker to define processing pipelines via a JSON configuration object, enabling operations such as overlaying images with transparency or rendering closed captions as text on video. The system supports real-time modifications, such as transforming an RTMP livestream by applying overlays or ingesting a hosted video via HLS to add subtitles before broadcasting or recording the output. Preview capabilities enable direct video input from sources like webcams or embedded devices, facilitating use cases such as AI analysis or combining multiple camera feeds into a composite view.
The technical implementation includes a Go-based HTTP server and Durable Object for low-level interface control, with an abstraction layer to decouple the controlling application from backend details. The controlling Worker can import @streamline/cloudflare and define a subclass of the Durable Object for application-specific logic, while local development bypasses authorization and uses a Docker instance with direct WebSocket connections for preview.