Scaling Camera File Processing at Netflix | by Netflix Technology Blog | Apr, 2026

Orchestrating Media Workflows By way of Strategic Collaboration

Authors: Eric Reinecke, Bhanu Srikanth

Introduction to Content material Hub’s Media Manufacturing Suite

At Netflix, we need to present filmmakers with the instruments they should produce content material at a world scale, with fast turnaround and selection from a unprecedented number of cameras, codecs, workflows, and collaborators. Each collection or movie arrives with its personal inventive ambitions and technical necessities. To cut back friction and hold productions shifting easily, we constructed Netflix’s Media Manufacturing Suite (MPS) with the aim of automating repeatable duties, standardizing key workflows, and giving productions extra time to give attention to inventive collaboration and craftsmanship.

A important a part of this effort is how we deal with picture processing and digicam metadata throughout the lots of of hours and terabytes of digicam footage that Netflix productions ingest every day. Relatively than construct each element from scratch, we selected to associate the place it made sense–particularly in areas the place the business already had trusted, battle-tested options.

This text explores how Netflix’s Media Manufacturing Suite integrates with FilmLight’s API (FLAPI) because the core studio media processing engine in Netflix’s cloud compute infrastructure, and the way that collaboration helps us ship smarter, extra dependable workflows at scale.

Why We Constructed MPS

As Netflix’s manufacturing slate grew, so did the complexity of file-based workflows. We noticed recurring challenges throughout productions:

• File wrangling sapping time from inventive decision-making
• Inconsistent media dealing with throughout exhibits, areas, or distributors
• Troublesome to audit handbook processes which are susceptible to human error
• Duplication of effort as groups reinvented related workflows for every manufacturing

Content material Hub Media Manufacturing Suite was created to deal with these ache factors. MPS is designed to:

• Convey effectivity, consistency, and high quality management to world productions
• Streamline media administration and motion from manufacturing by post-production
• Scale back time spent on non-creative file administration
• Decrease human error whereas maximizing inventive time

To realize this, MPS wanted a strong, versatile, and trusted technique to deal with camera-original media and metadata at scale.

The Proper Instrument for the Job

From the beginning, we knew that constructing a world-class picture processing engine in-house is a major, long-term dedication: one that will require deep, steady collaboration with digicam producers and the broader business.

When designing the system, we set out some core necessities:

• Examine, trim, and transcode unique digicam information and metadata for any Netflix manufacturing with trusted coloration science
• Help all kinds of cameras and recording codecs used worldwide whereas staying present as new ones are launched
• Run nicely in our paved-path encoding infrastructure, enabling us to benefit from confirmed compute and storage scalability with strong observability

FilmLight develops Baselight and Daylight, that are generally used within the business for coloration grading, dailies, and transcoding. Their FilmLight API (FLAPI) permits us to make use of that very same media processing engine as a backend API.

Relatively than duplicating that work, we selected to combine. FilmLight grew to become a trusted expertise associate, and FLAPI is now a foundational a part of how MPS processes media.

The Media Processing Engine

MPS shouldn’t be a single software; it’s an ecosystem of instruments and providers that assist Netflix productions globally. Inside that ecosystem, the FilmLight API performs the next key roles.

1. Parsing digicam metadata on ingest

Productions add media to Netflix’s Content material Hub with ASC MHL (Media Hash Checklist) information to make sure completeness and integrity of preliminary ingest, however quickly after, it’s essential to grasp the technical traits of every piece of media. We name this workflow section “inspection.”

Press enter or click on to view picture in full dimension

At this stage, we:

• Use FLAPI to collect digicam metadata from the unique digicam information
• Conform the workflow important fields to Netflix’s normalized schema
• Make it searchable and reusable for downstream processes

This metadata is integral to:

• Matching footage primarily based on timing and reel identify for automated retrieval
• Debugging (e.g., why a shot seems a sure method after processing)
• Validations and checks throughout the pipeline

FLAPI supplies constant, camera-aware perception into footage which will have originated anyplace on the planet. Moreover, since we’re in a position to bundle FLAPI in a Docker picture, we will deploy nearly equivalent code to each cloud and our manufacturing compute and storage facilities around the globe, making certain a constant evaluation of footage wherever it could exist.

2. Producing VFX plates and different deliverables

Visible results workflows always push picture processing pipelines to their absolute limits. For MPS to succeed, it should generate pictures with correct framing, constant coloration administration, and appropriate debayering/decoding parameters — all whereas sustaining fast turnaround occasions.

To realize this, we leverage Netflix’s Cosmos compute and storage platform and use open requirements to supply predictable and constant inventive management.

At this section, we use the FilmLight API to:

• Debayer unique digicam information with the proper format-specific decoding parameters
• Crop and de-squeeze pictures utilizing Framing Determination Lists (ASC FDL) to make sure spatial inventive choices are preserved
• Apply ACES Metadata Recordsdata (AMF), offering repeatable coloration pipelines from dailies by ending
• Generate an array of media deliverables in various codecs

These processes are automated, repeatable, and auditable. We ship AMFs alongside the OpenEXRs to make sure recipients know precisely what coloration transforms are already utilized, and which have to be utilized to match dailies.

As a result of we use FilmLight’s instruments on the backend, our workflow specialists can use Baselight on their workstations to manually validate pipeline choices for productions earlier than the primary day of principal pictures.

The Media Processing Manufacturing unit within the Cloud

Discovering an engine that competently processes media consistent with open requirements is a crucial a part of the equation. To maximise affect, we need to make these instruments obtainable to all the filmmakers we work with. Fortunately, we’re no strangers to scaled processing at Netflix, and our Cosmos compute platform was prepared for the job!

Cloud-first integration

The standard mannequin for this sort of processing in filmmaking has been to put money into beefy computer systems with giant GPUs and high-performance storage arrays to tear by debayering and encoding at breakneck velocity. Nevertheless, constraints within the cloud surroundings are completely different.

Elements which are important for instruments in our runtime surroundings embody that they:

• Are packageable as Serverless Capabilities in Linux Docker pictures that may be shortly invoked to run a single unit of labor and shut down on completion
• Can run on CPU-only cases to permit us to benefit from a big selection of accessible compute
• Help headless invocation through Java, Python, or CLI
• Function statelessly, so when issues do go mistaken, we will merely terminate and re-launch the employee

Working inside these constraints lets us give attention to rising throughput through parallel encoding moderately than specializing in single-instance processing energy. We are able to then goal the candy spot of the associated fee/efficiency effectivity curve whereas nonetheless hitting our goal turnaround occasions.

When instruments are API-driven, simply packaged in Linux containers, and don’t require a variety of exterior state administration, Netflix can shortly combine and deploy them with operational reliability. FilmLight API match the invoice for us. At Netflix, we leverage:

• Java and Python as the first integration languages
• Ubuntu-based Docker pictures with Java and Python code to show performance to our workflows
• CPU cases within the cloud and native compute facilities for working inspection, rendering, and trimming jobs

Whereas FLAPI additionally helps GPU rendering, CPU cases give us entry to a a lot wider phase of Netflix’s huge encoding compute pool and liberate GPU cases for different workloads.

To make use of FilmLight API, we bundle it in a bundle that may be simply put in through a Dockerfile. Then, we constructed Cosmos Stratum Capabilities that settle for an enter clip, output location, and ranging parameters comparable to body ranges and AMF or FDL information when debayering footage. These features might be shortly invoked to course of a single clip or sub-segment of a clip and shut down once more to liberate assets.

Elastic scaling for manufacturing workloads

Manufacturing workloads are inherently spiky:

• A quiet day on set could imply minimal new footage to examine.
• A full VFX turnover or pulling trimmed OCF for ending would possibly require hundreds of parallel renders in a short while window.

By deploying FLAPI within the cloud as features, MPS can:

• Allocate compute on demand and launch it when our work queue dies down
• Keep away from tying capability to a set pool of native {hardware}
• Easy demand throughout many forms of encoding workload in a shared useful resource pool

This elasticity lets us swarm pull requests to get them by shortly, then instantly yield assets again to decrease precedence workloads. Even in peak manufacturing intervals, we keep away from the ache of manually managing render queues and prioritization by avoiding fastened useful resource allocation. All this implies lightning-fast turnaround occasions and much less anxiousness round deadlines for our filmmakers.

Designed for Seasoned Execs and Rising Filmmakers

Netflix productions vary from extremely skilled groups with very particular workflows to newer groups who could also be much less acquainted with potential pitfalls in advanced file-based pipelines.

MPS is designed to assist each:

• Trade veterans who have to configure exact, bespoke workflows and belief that underlying picture processing will respect these choices.
• Productions with out a coloration scientist on workers — those that profit from guardrails and sane defaults that assist them keep away from widespread workflow points (e.g., mismatched coloration transforms, inconsistent debayering, or incomplete metadata dealing with).

The partnership with FilmLight lets Netflix give attention to workflow design, orchestration, and manufacturing assist, whereas FilmLight focuses on offering competent dealing with of all kinds of digicam codecs with world-class picture science!

Collaboration and Co-Evolution

Netflix aimed to combine MPS right into a wider instrument ecosystem by creating a complete resolution primarily based on rising open requirements, moderately than making MPS a self-contained system. Integrating FLAPI into our system requires greater than an API reference–it requires ongoing partnership. FilmLight labored intently with Netflix groups to:

• Align on function roadmaps, notably round new digicam codecs and open requirements
• Validate the accuracy and efficiency of key operations
• Debug edge circumstances found in large-scale, real-world workloads
• Evolve the API in ways in which serve each Netflix and the broader business
• Create a constructive suggestions cycle with open requirements like ACES and ASC FDL to unravel for gaps when the rubber hits the highway

One instance of this has been with the implementation of ACES 2. FilmLight’s builders shortly offered a roadmap for assist. As our engineering groups collaborated on integration, we additionally offered suggestions to the ACES technical management to shortly handle integration challenges and take a look at drive updates in our pipeline.

This collaborative relationship–constructed on open communication, joint validation, and suggestions to the larger business–is how we routinely work with FilmLight to make sure we’re not simply constructing one thing that works for our exhibits, but additionally driving a wholesome tooling and requirements ecosystem.

Influence

Whereas a lot of this work takes place behind the scenes, its affect is felt instantly by our productions. Our aim with constructing MPS is for producers, publish supervisors, and distributors to expertise:

• Fewer delays brought on by lacking, incomplete, or incorrect media
• Quicker turnaround on VFX plates and different technical deliverables
• Extra predictable, constant handoffs between editorial, coloration, and VFX
• Much less time spent troubleshooting technical points, and extra time centered on inventive overview

In follow, this typically exhibits up because the absence of disaster: the time a VFX vendor doesn’t must request a re-delivery, or the time editorial doesn’t have to attend for corrected plates, or the time the colour facility doesn’t must reinvent a tone-mapping path as a result of the AMF and ACES pipeline are already in place.

Trying Forward

As digicam expertise, codecs, open requirements, and manufacturing workflows proceed to evolve, so will MPS. The guiding ideas stay:

• Automate what’s repeatable
• Centralize what advantages from standardization
• Companion the place deep area experience already exists

The mixing with FilmLight API is one instance of this philosophy in motion. By treating picture processing as a specialised self-discipline and collaborating with a trusted business associate, Netflix is delivering smarter, extra dependable workflows to productions worldwide.

At its core, this partnership helps a easy aim: scale back handbook workflow and power administration, giving filmmakers extra time to inform tales.

Acknowledgements

This mission is the results of collaboration and iteration over a few years. Along with the authors, the next folks have contributed to this work:

• Matthew Donato
• Prabh Nallani
• Andy Schuler
• Jesse Korosi