Skip to main content
Slurpee3 is Pixwel’s media-processing engine. When a file is ingested, an encode is requested, or an automated localization order runs, Slurpee3 does the actual media work — probing, transcoding, rendering subtitles and graphics, and writing the resulting files back to storage. It’s the serverless successor to the original Slurpee (and Slurpee2), which were long-running PHP/SWF importers. Slurpee3 replaces that with AWS Step Functions orchestrating Lambda functions, plus AWS Elemental MediaConvert for transcoding and AWS Batch for subtitle/graphics rendering.

​
Architecture

PieceRole
Step FunctionsState machines that orchestrate each pipeline, with waitForTaskToken pauses for external verification.
LambdaThe individual steps (mediainfo, ingestPrepare, gfxEncode, createFile, …). Mostly Node.js, some Python.
MediaConvertAWS Elemental MediaConvert does the transcoding (GfxEncode, the encode workflow).
AWS BatchContainerized subtitle/graphics rendering (GfxSubtitles, a batch:submitJob.sync task).
S3 bucketsIngestOutputBucket, EncodeOutputBucket, GfxEncodeOutputBucket — outputs land here; ObjectCreated events trigger the completion Lambdas.
EventBridgeRules that fire completion handlers (e.g. EventGfxEncodeAudioComplete).
It’s defined as AWS SAM (slurpee3/template.yaml, deployed via deploy.sh) alongside a CDK project (slurpee3/cdk/, including the autosub stack). Source lives under slurpee3/workflows/{ingest,encodes,localization}.

​
The pipelines

​
Ingest

Runs when a source file is uploaded (see Ingests for the product view):
  1. mediainfo — read media metadata (always).
  2. ingestScratchPreview — a quick scratch preview (video only).
  3. ingestPrepare — ffprobe for video; updates the ingest’s status, mediainfo, selected, and ffprobe data.
  4. Wait for verification — a waitForTaskToken pause while the detected metadata is confirmed.
  5. ingestProcess — generate the preview and thumbnail (video only).
  6. copyFile / createFile — copy the download file when a download is requested.
  7. ingestVerify — finalize, updating the asset and ingest records.
ingestError propagates failures to the ingest status.

​
Encodes

Produces deliverable encodes through MediaConvert: createFile sets up the output, encode submits the MediaConvert job, and encodeComplete (triggered by an S3 ObjectCreated event on EncodeOutputBucket) registers the finished file back on the platform.

​
Localization (gfx)

The state machine in statemachine/gfx.asl.json — “Generates gfx render for a submitted order master file” — is what powers Automated Localization (subtitles and graphics):
  1. GfxPrepare — set up the render from the order’s master and translation.
  2. GfxEncode — MediaConvert transcode of the base video.
  3. GfxEncodeAudio — audio handling (waitForTaskToken).
  4. GfxSubtitles — AWS Batch container that renders/burns subtitles and graphics.
  5. GfxEncodeOffline — produce the reviewable offline.
  6. CopyFiles — a map state fanning out CopyFile to place outputs.
  7. GfxCreateOffline — register the offline back on the work request.
GfxError handles failures, and S3-triggered completion Lambdas (gfxEncodeComplete, gfxEncodeOfflineComplete, gfxEncodeAudioComplete) advance the flow as each render lands.

​
How it connects to platform2

The API starts the work — an ingest or a work request kicks off the relevant state machine. Slurpee3 processes the media and calls back to the API to update ingest, asset, and order status as steps complete (the waitForTaskToken pattern lets the API or a verification step resume a paused execution). Finished files land in S3 and are registered as platform files.

​
Supporting pieces

  • mediainfo/ — a Lambda wrapping the MediaInfo CLI (bundles libcurl) for metadata extraction.
  • copyFile/ — a Python Lambda for S3-to-S3 copies.
  • collect-metrics/ — gathers pipeline metrics.
  • cdk/autosub/ — the CDK stack for the autosub side of localization.

​
Deploying

  • SAM: cd slurpee3 && ./deploy.sh (uses template.yaml / packaged.yml).
  • CDK: from slurpee3/cdk/, the standard cdk deploy flow.
StepFunctions bills per transition, so the workflows keep state machines lean — choices are pre-computed into simple string/boolean hinges, and a step is only crossed when it genuinely needs orchestration (e.g. spanning Lambda → Batch → MediaConvert).