May 18, 2024

Liwei Guo, Vinicius Carvalho, Anush Moorthy, Aditya Mavlankar, Lishan Zhu

That is the second publish in a multi-part sequence from Netflix. See right here for Half 1 which supplies an summary of our efforts in rebuilding the Netflix video processing pipeline with microservices. This weblog dives into the main points of constructing our Video Encoding Service (VES), and shares our learnings.

Cosmos is the following technology media computing platform at Netflix. Combining microservice structure with asynchronous workflows and serverless features, Cosmos goals to modernize Netflix’s media processing pipelines with improved flexibility, effectivity, and developer productiveness. Previously few years, the video staff inside Encoding Applied sciences (ET) has been engaged on rebuilding your complete video pipeline on Cosmos.

This new pipeline consists of quite a lot of microservices, every devoted to a single performance. One such microservice is Video Encoding Service (VES). Encoding is a vital part of the video pipeline. At a excessive degree, it takes an ingested mezzanine and encodes it right into a video stream that’s appropriate for Netflix streaming or serves some studio/manufacturing use case. Within the case of Netflix, there are a selection of necessities for this service:

  • Given the big selection of units from cell phones to browsers to Good TVs, a number of codec codecs, resolutions, and high quality ranges must be supported.
  • Chunked encoding is a should to satisfy the latency necessities of our enterprise wants, and use instances with totally different ranges of latency sensitivity must be accommodated.
  • The potential of steady launch is essential for enabling quick product innovation in each streaming and studio areas.
  • There’s a large quantity of encoding jobs day by day. The service must be cost-efficient and take advantage of use of accessible assets.

On this tech weblog, we’ll stroll via how we constructed VES to attain the above targets and can share quite a lot of classes we discovered from constructing microservices. Please observe that for simplicity, we’ve got chosen to omit sure Netflix-specific particulars that aren’t integral to the first message of this weblog publish.

A Cosmos microservice consists of three layers: an API layer (Optimus) that takes in requests, a workflow layer (Plato) that orchestrates the media processing flows, and a serverless computing layer (Stratum) that processes the media. These three layers talk asynchronously via a home-grown, priority-based messaging system known as Timestone. We selected Protobuf because the payload format for its excessive effectivity and mature cross-platform help.

To assist service builders get a head begin, the Cosmos platform supplies a robust service generator. This generator options an intuitive UI. With just a few clicks, it creates a primary but full Cosmos service: code repositories for all 3 layers are created; all platform capabilities, together with discovery, logging, tracing, and many others., are enabled; launch pipelines are arrange and dashboards are readily accessible. We are able to instantly begin including video encoding logic and deploy the service to the cloud for experimentation.


Because the API layer, Optimus serves because the gateway into VES, which means service customers can solely work together with VES via Optimus. The outlined API interface is a powerful contract between VES and the exterior world. So long as the API is secure, customers are shielded from inside modifications in VES. This decoupling is instrumental in enabling quicker iterations of VES internals.

As a single-purpose service, the API of VES is sort of clear. We outlined an endpoint encodeVideo that takes an EncodeRequest and returns an EncodeResponse (in an async manner via Timestone messages). The EncodeRequest object accommodates details about the supply video in addition to the encoding recipe. All the necessities of the encoded video (codec, decision, and many others.) in addition to the controls for latency (chunking directives) are uncovered via the info mannequin of the encoding recipe.

//protobuf definition 

message EncodeRequest
VideoSource video_source = 1;//supply to be encoded
Recipe recipe = 2; //together with encoding format, decision, and many others.

message EncodeResponse
OutputVideo output_video = 1; //encoded video
Error error = 2; //error message (elective)

message Recipe
Codec codec = 1; //together with codec format, profile, degree, and many others.
Decision decision = 2;
ChunkingDirectives chunking_directives = 3;

Like another Cosmos service, the platform mechanically generates an RPC consumer primarily based on the VES API knowledge mannequin, which customers can use to construct the request and invoke VES. As soon as an incoming request is obtained, Optimus performs validations, and (when relevant) converts the incoming knowledge into an inside knowledge mannequin earlier than passing it to the following layer, Plato.

Like another Cosmos service, the platform mechanically generates an RPC consumer primarily based on the VES API knowledge mannequin, which customers can use to construct the request and invoke VES. As soon as an incoming request is obtained, Optimus performs validations, and (when relevant) converts the incoming knowledge into an inside knowledge mannequin earlier than passing it to the following layer, Plato.

The workflow layer, Plato, governs the media processing steps. The Cosmos platform helps two programming paradigms for Plato: ahead chaining rule engine and Directed Acyclic Graph (DAG). VES has a linear workflow, so we selected DAG for its simplicity.

In a DAG, the workflow is represented by nodes and edges. Nodes signify levels within the workflow, whereas edges signify dependencies — a stage is simply able to execute when all its dependencies have been accomplished. VES requires parallel encoding of video chunks to satisfy its latency and resilience targets. This workflow-level parallelism is facilitated by the DAG via a MapReduce mode. Nodes may be annotated to point this relationship, and a Cut back node will solely be triggered when all its related Map nodes are prepared.

For the VES workflow, we outlined 5 Nodes and their related edges, that are visualized within the following graph:

  • Splitter Node: This node divides the video into chunks primarily based on the chunking directives within the recipe.
  • Encoder Node: This node encodes a video chunk. It’s a Map node.
  • Assembler Node: This node stitches the encoded chunks collectively. It’s a Cut back node.
  • Validator Node: This node performs the validation of the encoded video.
  • Notifier Node: This node notifies the API layer as soon as your complete workflow is accomplished.

On this workflow, nodes such because the Notifier carry out very light-weight operations and may be immediately executed within the Plato runtime. Nonetheless, resource-intensive operations must be delegated to the computing layer (Stratum), or one other service. Plato invokes Stratum features for duties akin to encoding and assembling, the place the nodes (Encoder and Assembler) publish messages to the corresponding message queues. The Validator node calls one other Cosmos service, the Video Validation Service, to validate the assembled encoded video.


The computing layer, Stratum, is the place media samples may be accessed. Builders of Cosmos providers create Stratum Capabilities to course of the media. They will deliver their very own media processing instruments, that are packaged into Docker photos of the Capabilities. These Docker photos are then revealed to our inside Docker registry, a part of Titus. In manufacturing, Titus mechanically scales situations primarily based on the depths of job queues.

VES must help encoding supply movies into quite a lot of codec codecs, together with AVC, AV1, and VP9, to call just a few. We use totally different encoder binaries (referred to easily as “encoders”) for various codec codecs. For AVC, a format that’s now 20 years previous, the encoder is sort of secure. Alternatively, the most recent addition to Netflix streaming, AV1, is constantly going via lively enhancements and experimentations, necessitating extra frequent encoder upgrades. ​​To successfully handle this variability, we determined to create a number of Stratum Capabilities, every devoted to a particular codec format and may be launched independently. This method ensures that upgrading one encoder won’t impression the VES service for different codec codecs, sustaining stability and efficiency throughout the board.

Inside the Stratum Perform, the Cosmos platform supplies abstractions for widespread media entry patterns. No matter file codecs, sources are uniformly introduced as regionally mounted frames. Equally, for output that must be endured within the cloud, the platform presents the method as writing to a neighborhood file. All particulars, akin to streaming of bytes and retrying on errors, are abstracted away. With the platform caring for the complexity of the infrastructure, the important code for video encoding within the Stratum Perform may very well be so simple as follows.

ffmpeg -i enter/supplypercent08d.j2k -vf ... -c:v libx264 ... output/encoding.264

Encoding is a resource-intensive course of, and the assets required are intently associated to the codec format and the encoding recipe. We performed benchmarking to grasp the useful resource utilization sample, notably CPU and RAM, for various encoding recipes. Primarily based on the outcomes, we leveraged the “container shaping” characteristic from the Cosmos platform.

We outlined quite a lot of totally different “container shapes”, specifying the allocations of assets like CPU and RAM.

# an instance definition of container form
group: containerShapeExample1
numCpus: 2
memoryInMB: 4000
networkInMbp: 750
diskSizeInMB: 12000

Routing guidelines are created to assign encoding jobs to totally different shapes primarily based on the mixture of codec format and encoding decision. This helps the platform carry out “bin packing”, thereby maximizing useful resource utilization.

An instance of “bin-packing”. The circles signify CPU cores and the realm represents the RAM. This 16-core EC2 occasion is filled with 5 encoding containers (rectangles) of three totally different shapes (indicated by totally different colours).

After we accomplished the event and testing of all three layers, VES was launched in manufacturing. Nonetheless, this didn’t mark the top of our work. Fairly the opposite, we believed and nonetheless do {that a} vital a part of a service’s worth is realized via iterations: supporting new enterprise wants, enhancing efficiency, and bettering resilience. An essential piece of our imaginative and prescient was for Cosmos providers to have the power to constantly launch code modifications to manufacturing in a secure method.

Specializing in a single performance, code modifications pertaining to a single characteristic addition in VES are usually small and cohesive, making them simple to evaluation. Since callers can solely work together with VES via its API, inside code is actually “implementation particulars” which are secure to vary. The specific API contract limits the check floor of VES. Moreover, the Cosmos platform supplies a pyramid-based testing framework to information builders in creating assessments at totally different ranges.

After testing and code evaluation, modifications are merged and are prepared for launch. The discharge pipeline is absolutely automated: after the merge, the pipeline checks out code, compiles, builds, runs unit/integration/end-to-end assessments as prescribed, and proceeds to full deployment if no points are encountered. Sometimes, it takes round half-hour from code merge to characteristic touchdown (a course of that took 2–4 weeks in our earlier technology platform!). The brief launch cycle supplies quicker suggestions to builders and helps them make needed updates whereas the context remains to be recent.

Screenshot of a launch pipeline run in our manufacturing surroundings

When working in manufacturing, the service continuously emits metrics and logs. They’re collected by the platform to visualise dashboards and to drive monitoring/alerting methods. Metrics deviating an excessive amount of from the baseline will set off alerts and might result in computerized service rollback (when the “canary” characteristic is enabled).

VES was the very first microservice that our staff constructed. We began with primary data of microservices and discovered a large number of classes alongside the way in which. These learnings deepened our understanding of microservices and have helped us enhance our design selections and choices.

Outline a Correct Service Scope

A precept of microservice structure is {that a} service must be constructed for a single performance. This sounds simple, however what precisely qualifies a “single performance”? “Encoding video” sounds good however wouldn’t “encode video into the AVC format” be an much more particular single-functionality?

After we began constructing the VES, we took the method of making a separate encoding service for every codec format. Whereas this has benefits akin to decoupled workflows, shortly we have been overwhelmed by the event overhead. Think about {that a} person requested us so as to add the watermarking functionality to the encoding. We wanted to make modifications to a number of microservices. What’s worse, modifications in all these providers are very related and primarily we’re including the identical code (and assessments) repeatedly. Such type of repetitive work can simply put on out builders.

The service introduced on this weblog is our second iteration of VES (sure, we already went via one iteration). On this model, we consolidated encodings for various codec codecs right into a single service. They share the identical API and workflow, whereas every codec format has its personal Stratum Capabilities. To this point this appears to strike an excellent stability: the widespread API and workflow reduces code repetition, whereas separate Stratum Capabilities assure impartial evolution of every codec format.

The modifications we made should not irreversible. If sometime sooner or later, the encoding of 1 specific codec format evolves into a very totally different workflow, we’ve got the choice to spin it off into its personal microservice.

Be Pragmatic about Knowledge Modeling

At first, we have been very strict about knowledge mannequin separation — we had a powerful perception that sharing equates to coupling, and coupling might result in potential disasters sooner or later. To keep away from this, for every service in addition to the three layers inside a service, we outlined its personal knowledge mannequin and constructed converters to translate between totally different knowledge fashions.

We ended up creating a number of knowledge fashions for points akin to bit-depth and determination throughout our system. To be truthful, this does have some deserves. For instance, our encoding pipeline helps totally different bit-depths for AVC encoding (8-bit) and AV1 encoding (10-bit). By defining each AVC.BitDepth and AV1.BitDepth, constraints on the bit-depth may be constructed into the info fashions. Nonetheless, it’s debatable whether or not the advantages of this differentiation energy outweigh the downsides, specifically a number of knowledge mannequin translations.

Ultimately, we created a library to host knowledge fashions for widespread ideas within the video area. Examples of such ideas embody body price, scan kind, colour area, and many others. As you may see, they’re extraordinarily widespread and secure. This “widespread” knowledge mannequin library is shared throughout all providers owned by the video staff, avoiding pointless duplications and knowledge conversions. Inside every service, extra knowledge fashions are outlined for service-specific objects.

Embrace Service API Adjustments

This will likely sound contradictory. Now we have been saying that an API is a powerful contract between the service and its customers, and holding an API secure shields customers from inside modifications. That is completely true. Nonetheless, none of us had a crystal ball once we have been designing the very first model of the service API. It’s inevitable that at a sure level, this API turns into insufficient. If we maintain the assumption that “the API can not change” too dearly, builders can be pressured to search out workarounds, that are virtually definitely sub-optimal.

There are numerous nice tech articles about gracefully evolving API. We imagine we even have a singular benefit: VES is a service inside to Netflix Encoding Applied sciences (ET). Our two customers, the Streaming Workflow Orchestrator and the Studio Workflow Orchestrator, are owned by the workflow staff inside ET. Our groups share the identical contexts and work in direction of widespread targets. If we imagine updating API is in the most effective curiosity of Netflix, we meet with them to hunt alignment. As soon as a consensus to replace the API is reached, groups collaborate to make sure a clean transition.

That is the second a part of our tech weblog sequence Rebuilding Netflix Video Pipeline with Microservices. On this publish, we described the constructing means of the Video Encoding Service (VES) intimately in addition to our learnings. Our pipeline features a few different providers that we plan to share about as effectively. Keep tuned for our future blogs on this subject of microservices!