How we reduced costs by switching from AWS MediaConvert to a Golang service

This article will focus mainly on the golang side of work, so if you're interested please continue reading

---

The problem:
I'm working at a real state company. We allow our users to upload videos showing their proprieties and we used AWS MediaConvert to handle compressing and transcoding but things got out of hand. With MediaConvert a single minute of video encoding costs $0.03 for each video quality at 720p and above and $0.015 for each video quality below 720p. That way, if you encode at 720p, 480p, and 360p you pay – $0.06 per minute which is $3.6 per hour.

A screenshot from AWS MediaConvert pricing page

When you don't have many videos and the videos themselves aren't really long this is fine, but it'll start to cost more when your user base grows.

---

The solution:
I used golang and ffmpeg to create a service that allow us to replace MediaConvert. Go allows you to write concurrent code and it's really fast compared to other high programming languages like Python, Java, etc. With this in hand you can build really powerful things.
I've used go-fluent-ffmpeg for ffmpeg integration with go - I'm planning to switch to cgo bindings for more performance!. I've got this idea from a former college of mine, a really talented engineer phr3nzy

Here's the flow for how things were done

• S3 is AWS simple storage service.
• Lambda is a serverless, event-driven compute service.
• SQS is a managed message queuing service.

---

Here's how the lambda code looks like

const aws = require("aws-sdk");
const sqs = new aws.SQS({ apiVersion: "2012-11-05" });
const s3 = new aws.S3({ apiVersion: "2006-03-01" });

exports.handler = async (event, context) => {
  const bucket = event.Records[0].s3.bucket.name;
  const key = decodeURIComponent(
    event.Records[0].s3.object.key.replace(/\+/g, " ")
  );

  try {
    const { ContentType, ContentLength } = await s3
      .headObject({ Bucket: bucket, Key: key })
      .promise();

    const [type] = ContentType.split("/"); // ["image", "jpeg"], ["video", "mp4"], ...

    if (type === "video") {
      // convert ContentLength from bytes to megabytes
      const size = ContentLength / (1024 * 2024);
      const messageBody = {
        originBucketName: bucket,
        originalFilePath: key,
        destinationBucketName: "",
        destinationBucketFolder: "",
        orientation: "landscape", // "landscape" || "portrait"
        resolution: "480", // 360 || 480 || 720
        size,
      };

      await sqs
        .sendMessage({
          MessageBody: JSON.stringify(messageBody),

          QueueUrl: process.ENV.VIDEO_QUEUE_URL,
        })
        .promise();
    }
  } catch (err) {
    console.log(err);
    throw new Error(err);
  }
};

---

After capturing the video data we send it to SQS then the go service should poll and process the video. I used Go AWS SDK to interact with AWS. and for the actual processing I've implemented a pipeline with three stages to process the videos.

---

Stage 1 - Downloading

func download(s3 *s3.Client, messages ...config.Message) <-chan config.DownloadedFile {
    out := make(chan config.DownloadedFile)
    go func() {
        defer close(out)
        for _, msg := range messages {
            fullpath, err := storage.DownloadObject(s3, msg.OriginBucketName, msg.Filename)
            if err != nil {
                log.Error(err.Error())
                return
            }
            name := strings.Split(msg.Filename, ".")
            format := strings.ToLower(name[len(name)-1])
            path := config.DownloadedFile{Fullpath: fullpath,
                Filename:                msg.Filename,
                DestinationBucketName:   msg.DestinationBucketName,
                DestinationBucketFolder: msg.DestinationBucketFolder,
                Resolution:              msg.Resolution,
                Orientation:             msg.Orientation,
                Format:                  format,
            }
            out <- path
        }
    }()
    return out
}

---

Stage 2 - Transcoding

func transcode(paths <-chan config.DownloadedFile) <-chan config.TranscodedFile {
    out := make(chan config.TranscodedFile)
    go func() {
        defer close(out)
        for file := range paths {
            unique := fmt.Sprintf("%s_%s_%s", file.Orientation, file.Resolution, file.Filename)
            outputPath := fmt.Sprintf("%s/%s", config.RootVideosDir, unique)
            ffmpeg := fluentffmpeg.NewCommand("")
            vError := ffmpeg.
                InputPath(file.Fullpath).
                FromFormat(file.Format).
                VideoBitRate(4 * 1042).
                FrameRate(30).
                AspectRatio(AspectRatios[file.Orientation]).
                Resolution(VideoResolutions[file.Resolution]).
                VideoCodec("libx265").
                ConstantRateFactor(VideoQualities[file.Resolution]).
                Preset("ultrafast").
                OutputFormat(file.Format).
                OutputPath(outputPath).
                Overwrite(true).
                Run()

            if vError != nil {
                log.Error(vError.Error())
                return
            }
            transcoded := config.TranscodedFile{
                DestinationBucketName:   file.DestinationBucketName,
                DestinationBucketFolder: file.DestinationBucketFolder,
                Fullpath:                file.Fullpath,
                LocalDiskPath:           outputPath,
                Orientation:             file.Orientation,
                Filename:                file.Filename}
            out <- transcoded
        }
    }()

    return out
}

---

Stage 3 - uploading

func upload(s3 *s3.Client, videos <-chan config.TranscodedFile) {
    go func() {
        defer close(errc)
        for video := range videos {
            dest := fmt.Sprintf("%s/%s",video.DestinationBucketFolder, video.Filename)
            storage.UploadVideo(s3, video.Fullpath, video.DestinationBucketName, dest)
            deleteFiles(video.Fullpath, video.LocalDiskPath)
        }
    }()
}

---

Finally I'd run the pipeline with this line

upload(s3Client,transcode(download(s3Client, messages...)))

Of course may need to add more but this is a showcase.

The code above uses goroutines which is how you write concurrent code in go.

Thanks for sticking to the end of the article!. Hope it helps.

Comments

[luan vu]

I think some is confusing here. The main progress is not your Go lang code but FFmpeg. You are also running on Lambda, so concurrent could be a dangerous thing. Furthermore, if your Lambda instances need several minutes, could be a huge impact on the cost.
Beware: You are using Ultrafast preset, that means the data loss is highest.

[Mustafa Balila]

No, I've deployed the service on a compute optimized EC2 - costs around $250 a month.
Isn't ultrafast for faster compression but bigger sizes? I did it because the size does't difference is't relevant compared to the speed

[luan vu]

My mistake. That's correct. Preset relevant to compression, not quality.
Use EC2 is good choice.

[robzer]

I've built something similar but using ECS Fargate running a linux container and with a simple bash script I was able to read the source video file from S3, process it using ffmpeg CLI and then upload the output on another S3 bucket

I've done a few tests manually and it's working perfectly fine, but I'm not sure how robust it is on a bigger scale

Did you considered this option too ?

[Paul Arah]

I'm a bit confused. If you're using an EC2 instance, what does your lambda function do?

[Renato Santos]

gets event notifications from s3 and put on sqs to enqueue to the go service

[Josh Dvir]

You can even drop the lambda function in favor of S3 event notification docs.aws.amazon.com/AmazonS3/lates...

[Mustafa Balila]

Didn't know about that, it's brilliant. Thanks!

[Jack Bridger]

Awesome article! Really nice architecture. Do you have a rough idea of your cost savings?

[Glaucio Guerra]

Hi,

Did you create some comparative performance test between the go solution and aws service?

Thanks for your article by the way.

[Mustafa Balila]

Hello. Yeah the go service is slightly slower than MediaConvert that's why we're using SQS queue. You can still achieve the same speed if you used a bigger instance I was using g4ad.xlarge