How Remotion Lambda works

This document describes the procedure that gets executed when a Remotion Lambda video render is triggered.

note

This document explains the Lambda architecture from version 4.0.165 on.
Previously, Lambda functions did not use Response Streaming, and instead saved chunks to S3.

1

A single Lambda function is invoked using renderMediaOnLambda(), renderMediaOnLambdaAndStream(), or the CLI. This invocation is called the main function.

2

The main function visits the Serve URL that it is being passed in a headless browser.

3

The main function finds the composition based on the composition ID that was passed and runs the props resolution algorithm to determine the props that should be passed to the video as well as the metadata (such as the duration in frames).

4

Based on the determined duration of the video and the concurrency , the video is split into chunks. Normally, a renderer function is invoked for each chunk. From v4.0.517, you can set concurrency to 1 to render the video on the main function.

5

Separate renderer functions use AWS Lambda Response Streaming to report progress and send the binary video chunks to the main function. This streaming step is skipped when the main function renders directly.

6

The main function concatenates the progress reports into a concise progress.json file and periodically uploads it to S3.

7

From v4.0.542, renderMediaOnLambdaAndStream() and the CLI receive progress updates streamed from the main function. getRenderProgress() remains available to check a render started separately.

8

As soon as all chunks have arrived in the main function, they get seamlessly concatenated. The concatenation algorithm is not a public API at the moment.

9

The main function uploads the final video to S3 and shuts down.

FAQ

Can I roll my own distributed renderer?

The seamless concatenation of chunks is not a public API at the moment.
You may render chunks using frameRange and audioCodec: "pcm-16" which you can concatenate using FFmpeg.

Building a distributed renderer is hard, and not recommended for most.

Will each chunk download all assets?

Each chunk will download all assets that are referenced in this chunk.
This can lead to assets being downloaded many times at the same time, which may overwhelm a server or trigger rate limits.
In addition, you pay for the bandwidth, even if the assets are on S3¹⁾.

Keep this in mind when designing your solution and consider using a CDN to serve assets.

---

¹⁾ An API for avoiding the S3 bandwidth charge is planned.