SVT Encore

SVT Encore is a scalable, open-source Transcoding-as-a-Service tool built on FFmpeg and Spring Boot. It provides a REST API to abstract the complexity of FFmpeg, making it suitable for integration into automated VOD (Video On Demand) transcoding pipelines.

Key Capabilities:

• Scalable Architecture: Decoupled web API and worker components using Redis-backed job queues.
• Profile-driven Transcoding: Uses YAML-based profiles with support for dynamic parameters via SpEL expressions.
• Advanced Features: Supports H.264, H.265, and any FFmpeg-supported codec; thumbnail generation (single and sprite sheets); VMAF quality metrics; and segmented (parallel chunk-based) transcoding.
• Monitoring: Provides HTTP progress callbacks for job monitoring.
• Security: Optional basic authentication with user and admin roles.
• API Exploration: Built-in Swagger UI for testing and exploring the REST API.

Encore operates as a distributed system consisting of two primary services that communicate through a Redis instance. This design allows you to scale the transcoding capacity horizontally by adding more workers.

Core Components:

• encore-web (REST API Service): The entry point for clients. It accepts transcoding jobs via REST, serves the Swagger UI, and can optionally process jobs itself. Clients poll this service to check job status.
• encore-worker (Headless Worker): Polls Redis for new jobs and executes the actual transcoding using FFmpeg. To increase throughput, you can run multiple encore-worker instances.
• Redis: Acts as the central data store for job data, the job queue, and the pub/sub mechanism used for job cancellation and segmented transcoding progress.
• FFmpeg: The underlying engine that performs the transcoding. It must be installed on any machine running encore-web (if polling is enabled) or encore-worker.

Jobs can contain multiple inputs, each assigned a unique videoLabel or audioLabel. To target a specific input in your transcoding profile, use the inputLabel field within an encode definition. This allows you to create separate encodes for different source streams (e.g., a main video stream and an alternative audio stream) within a single job.

// Job with multiple inputs
{
  "profile": "composite",
  "outputFolder": "/output",
  "baseName": "composite_output",
  "inputs": [
    {
      "type": "AudioVideo",
      "uri": "/input/main_video.mp4",
      "videoLabel": "main",
      "audioLabel": "main"
    },
    {
      "type": "Audio",
      "uri": "/input/alt_audio.wav",
      "audioLabel": "alt"
    }
  ]
}

# Profile referencing those labels
encodes:
  - type: AudioEncode
    bitrate: 128k
    suffix: _STEREO
    inputLabel: main

  - type: AudioEncode
    bitrate: 128k
    suffix: _STEREO_ALT
    inputLabel: alt

Audio mix presets define how audio channels are mapped when converting between channel layouts (e.g., 5.1 surround to stereo). When an AudioEncode performs a conversion, Encore resolves the pan filter in this specific order:

1. pan-mapping[source][target]: An explicit mapping for that specific source $\rightarrow$ target pair.
2. default-pan[target]: A default pan filter for the target layout. This is only used for downmixing (where the target has fewer or equal channels than the source).
3. fallback-to-auto: If true, fall back to FFmpeg's automatic channel mapping.
4. Error: If no match is found, the encode fails (or is skipped if optional: true).

Upmixing Note: Upmixing (e.g., stereo to 5.1) generally requires an explicit pan-mapping entry or relying on fallback-to-auto.

encore-settings:
  encoding:
    audio-mix-presets:
      default:
        fallback-to-auto: true
        default-pan:
          stereo: "FL=FL+0.707*FC+0.707*BL+0.707*SL|FR=FR+0.707*FC+0.707*BR+0.707*SR"
        pan-mapping:
          mono:
            stereo: "FL=0.707*FC|FR=0.707*FC"

A production deployment of Encore consists of three primary components:

1. encore-web: The REST API. It can run behind a load balancer for high availability. By default, it also polls for jobs to process them itself. To restrict it to only serving the API and offload all transcoding to workers, set encore-settings.poll-disabled=true.
2. encore-worker: A headless component that polls a single job from the queue, performs the transcoding, and then exits. You can scale horizontally by running multiple worker instances. To keep a worker polling until the queue is completely empty, set encore-settings.worker-drain-queue=true.
3. Redis 8+: Acts as the job storage and queue. It must have the JSON and Search modules enabled. For production, ensure persistence is enabled.

Encore uses priority queues to prevent long-running, low-priority transcoding jobs from blocking urgent, high-priority jobs.

Core Concepts

• Priority Classes: The encore-settings.concurrency setting defines the number of priority queues (classes) in Redis. Each class has its own queue and its own pool of workers.
• Job Routing: Jobs are routed to a queue based on their priority (0-100). Higher-priority jobs within a class are processed before lower-priority ones in that same class.
• Worker Behavior:
  • encore-worker uses a single thread and polls a single queue. You must pin a worker to a specific queue using encore-settings.poll-queue.
  • encore-web (with polling enabled) starts one thread per queue.

Recommended Production Pattern

For production, it is recommended to run encore-web with poll-disabled: true and dedicate specific encore-worker pods to each priority class to ensure isolation.

Priority-to-Queue Mapping Example

With concurrency: 2, the mapping is:

• Priority 50–100: Queue 0 (High priority)
• Priority 0–49: Queue 1 (Low priority)

Note: Queue 0 is always the highest-priority class, and queue concurrency - 1 is always the lowest.

graph LR
    API[encore-web<br/>poll-disabled: true] --> Redis[(Redis)]
    Redis --> HP[high-priority-worker<br/>poll-queue: 0]
    Redis --> HP2[high-priority-worker<br/>poll-queue: 0]
    Redis --> LP[low-priority-worker<br/>poll-queue: 1]
    Redis --> LP2[low-priority-worker<br/>poll-queue: 1]