David J. Berg, Romain Cledat, Kayla Seeley, Shashank Srikanth, Chaoying Wang, Darin Yu
Netflix makes use of knowledge science and machine studying throughout all aspects of the corporate, powering a variety of enterprise purposes from our inner infrastructure and content material demand modeling to media understanding. The Machine Learning Platform (MLP) crew at Netflix gives a whole ecosystem of instruments round Metaflow, an open supply machine studying infrastructure framework we began, to empower knowledge scientists and machine studying practitioners to construct and handle a wide range of ML programs.
Since its inception, Metaflow has been designed to offer a human-friendly API for constructing knowledge and ML (and right this moment AI) purposes and deploying them in our manufacturing infrastructure frictionlessly. While human-friendly APIs are pleasant, it’s actually the integrations to our manufacturing programs that give Metaflow its superpowers. Without these integrations, tasks could be caught on the prototyping stage, or they must be maintained as outliers outdoors the programs maintained by our engineering groups, incurring unsustainable operational overhead.
Given the very numerous set of ML and AI use circumstances we help — right this moment we have now tons of of Metaflow tasks deployed internally — we don’t count on all tasks to comply with the identical path from prototype to manufacturing. Instead, we offer a strong foundational layer with integrations to our company-wide knowledge, compute, and orchestration platform, in addition to numerous paths to deploy purposes to manufacturing easily. On high of this, groups have constructed their very own domain-specific libraries to help their particular use circumstances and desires.
In this text, we cowl a number of key integrations that we offer for numerous layers of the Metaflow stack at Netflix, as illustrated above. We may also showcase real-life ML tasks that depend on them, to present an concept of the breadth of tasks we help. Note that each one tasks leverage a number of integrations, however we spotlight them within the context of the combination that they use most prominently. Importantly, all of the use circumstances had been engineered by practitioners themselves.
These integrations are carried out by means of Metaflow’s extension mechanism which is publicly obtainable however topic to vary, and therefore not part of Metaflow’s secure API but. If you might be interested in implementing your personal extensions, get in contact with us on the Metaflow group Slack.
Let’s go over the stack layer by layer, beginning with probably the most foundational integrations.
Our principal knowledge lake is hosted on S3, organized as Apache Iceberg tables. For ETL and different heavy lifting of information, we primarily depend on Apache Spark. In addition to Spark, we need to help last-mile knowledge processing in Python, addressing use circumstances resembling function transformations, batch inference, and coaching. Occasionally, these use circumstances contain terabytes of information, so we have now to concentrate to efficiency.
To allow quick, scalable, and sturdy entry to the Netflix knowledge warehouse, we have now developed a Fast Data library for Metaflow, which leverages high-performance elements from the Python knowledge ecosystem:
As depicted within the diagram, the Fast Data library consists of two principal interfaces:
- The
Tableobject is answerable for interacting with the Netflix knowledge warehouse which incorporates parsing Iceberg (or legacy Hive) desk metadata, resolving partitions and Parquet information for studying. Recently, we added help for the write path, so tables may be up to date as nicely utilizing the library. - Once we have now found the Parquet information to be processed,
MetaflowDataBodytakes over: it downloads knowledge utilizing Metaflow’s high-throughput S3 consumer on to the method’ reminiscence, which usually outperforms studying of native information.
We use Apache Arrow to decode Parquet and to host an in-memory illustration of information. The consumer can select probably the most appropriate software for manipulating knowledge, resembling Pandas or Polars to make use of a dataframe API, or one in all our inner C++ libraries for numerous high-performance operations. Thanks to Arrow, knowledge may be accessed by means of these libraries in a zero-copy vogue.
We additionally take note of dependency points: (Py)Arrow is a dependency of many ML and knowledge libraries, so we don’t need our customized C++ extensions to depend upon a particular model of Arrow, which may simply result in unresolvable dependency graphs. Instead, within the type of nanoarrow, our Fast Data library solely depends on the secure Arrow C knowledge interface, producing a hermetically sealed library with no exterior dependencies.
Example use case: Content Knowledge Graph
Our data graph of the leisure world encodes relationships between titles, actors and different attributes of a movie or sequence, supporting all features of enterprise at Netflix.
A key problem in making a data graph is entity decision. There could also be many alternative representations of barely totally different or conflicting details about a title which should be resolved. This is often performed by means of a pairwise matching process for every entity which turns into non-trivial to do at scale.
This undertaking leverages Fast Data and horizontal scaling with Metaflow’s foreach assemble to load giant quantities of title data — roughly a billion pairs — saved within the Netflix Data Warehouse, so the pairs may be matched in parallel throughout many Metaflow duties.
We use metaflow.Table to resolve all enter shards that are distributed to Metaflow duties that are answerable for processing terabytes of information collectively. Each process hundreds the info utilizing metaflow.MetaflowDataBody, performs matching utilizing Pandas, and populates a corresponding shard in an output Table. Finally, when all matching is finished and knowledge is written the brand new desk is dedicated so it may be learn by different jobs.
Whereas open-source customers of Metaflow depend on AWS Batch or Kubernetes because the compute backend, we depend on our centralized compute-platform, Titus. Under the hood, Titus is powered by Kubernetes, however it gives a thick layer of enhancements over off-the-shelf Kubernetes, to make it extra observable, safe, scalable, and cost-efficient.
By focusing on @titus, Metaflow duties profit from these battle-hardened options out of the field, with no in-depth technical data or engineering required from the ML engineers or knowledge scientist finish. However, with the intention to profit from scalable compute, we have to assist the developer to bundle and rehydrate the entire execution surroundings of a undertaking in a distant pod in a reproducible method (ideally rapidly). Specifically, we don’t need to ask builders to handle Docker pictures of their very own manually, which rapidly leads to extra issues than it solves.
This is why Metaflow gives help for dependency administration out of the field. Originally, we supported solely @conda, however based mostly on our work on Portable Execution Environments, open-source Metaflow gained help for @pypi a number of months in the past as nicely.
Example use case: Building mannequin explainers
Here’s a captivating instance of the usefulness of moveable execution environments. For lots of our purposes, mannequin explainability issues. Stakeholders like to grasp why fashions produce a sure output and why their habits modifications over time.
There are a number of methods to offer explainability to fashions however a technique is to coach an explainer mannequin based mostly on every skilled mannequin. Without going into the main points of how that is performed precisely, suffice to say that Netflix trains loads of fashions, so we have to practice loads of explainers too.
Thanks to Metaflow, we are able to permit every utility to decide on the most effective modeling strategy for his or her use circumstances. Correspondingly, every utility brings its personal bespoke set of dependencies. Training an explainer mannequin subsequently requires:
- Access to the unique mannequin and its coaching surroundings, and
- Dependencies particular to constructing the explainer mannequin.
This poses an attention-grabbing problem in dependency administration: we want a higher-order coaching system, “Explainer flow” within the determine beneath, which is ready to take a full execution surroundings of one other coaching system as an enter and produce a mannequin based mostly on it.
Explainer circulate is event-triggered by an upstream circulate, such Model A, B, C flows within the illustration. The build_environment step makes use of the metaflow surroundings command offered by our moveable environments, to construct an surroundings that features each the necessities of the enter mannequin in addition to these wanted to construct the explainer mannequin itself.
The constructed surroundings is given a singular identify that relies on the run identifier (to offer uniqueness) in addition to the mannequin kind. Given this surroundings, the train_explainer step is then capable of seek advice from this uniquely named surroundings and function in an surroundings that may each entry the enter mannequin in addition to practice the explainer mannequin. Note that, not like in typical flows utilizing vanilla @conda or @pypi, the moveable environments extension permits customers to additionally fetch these environments straight at execution time versus at deploy time which subsequently permits customers to, as on this case, resolve the surroundings proper earlier than utilizing it within the subsequent step.
If knowledge is the gas of ML and the compute layer is the muscle, then the nerves should be the orchestration layer. We have talked in regards to the significance of a production-grade workflow orchestrator within the context of Metaflow when we launched help for AWS Step Functions years in the past. Since then, open-source Metaflow has gained help for Argo Workflows, a Kubernetes-native orchestrator, in addition to help for Airflow which remains to be extensively utilized by knowledge engineering groups.
Internally, we use a manufacturing workflow orchestrator referred to as Maestro. The Maestro publish shares particulars about how the system helps scalability, high-availability, and usefulness, which offer the spine for all of our Metaflow tasks in manufacturing.
A vastly necessary element that always goes neglected is event-triggering: it permits a crew to combine their Metaflow flows to surrounding programs upstream (e.g. ETL workflows), in addition to downstream (e.g. flows managed by different groups), utilizing a protocol shared by the entire group, as exemplified by the instance use case beneath.
Example use case: Content choice making
One of probably the most business-critical programs working on Metaflow helps our content material choice making, that’s, the query of what content material Netflix ought to deliver to the service. We help a large scale of over 260M subscribers spanning over 190 nations representing vastly numerous cultures and tastes, all of whom we need to delight with our content material slate. Reflecting the breadth and depth of the problem, the programs and fashions specializing in the query have grown to be very subtle.
We strategy the query from a number of angles however we have now a core set of information pipelines and fashions that present a basis for choice making. To illustrate the complexity of simply the core elements, think about this high-level diagram:
In this diagram, grey packing containers symbolize integrations to accomplice groups downstream and upstream, inexperienced packing containers are numerous ETL pipelines, and blue packing containers are Metaflow flows. These packing containers encapsulate tons of of superior fashions and complicated enterprise logic, dealing with huge quantities of information every day.
Despite its complexity, the system is managed by a comparatively small crew of engineers and knowledge scientists autonomously. This is made potential by a number of key options of Metaflow:
The crew has additionally developed their very own domain-specific libraries and configuration administration instruments, which assist them enhance and function the system.
To produce enterprise worth, all our Metaflow tasks are deployed to work with different manufacturing programs. In many circumstances, the combination is perhaps through shared tables in our knowledge warehouse. In different circumstances, it’s extra handy to share the outcomes through a low-latency API.
Notably, not all API-based deployments require real-time analysis, which we cowl within the part beneath. We have a variety of business-critical purposes the place some or all predictions may be precomputed, guaranteeing the bottom potential latency and operationally easy excessive availability on the international scale.
We have developed an formally supported sample to cowl such use circumstances. While the system depends on our inner caching infrastructure, you might comply with the identical sample utilizing providers like Amazon ElasticCache or DynamoDB.
Example use case: Content efficiency visualization
The historic efficiency of titles is utilized by choice makers to grasp and enhance the movie and sequence catalog. Performance metrics may be advanced and are sometimes finest understood by people with visualizations that break down the metrics throughout parameters of curiosity interactively. Content choice makers are geared up with self-serve visualizations by means of a real-time internet utility constructed with metaflow.Cache, which is accessed by means of an API supplied with metaflow.Hosting.
A every day scheduled Metaflow job computes mixture portions of curiosity in parallel. The job writes a big quantity of outcomes to a web-based key-value retailer utilizing metaflow.Cache. A Streamlit app homes the visualization software program and knowledge aggregation logic. Users can dynamically change parameters of the visualization utility and in real-time a message is distributed to a easy Metaflow internet hosting service which appears up values within the cache, performs computation, and returns the outcomes as a JSON blob to the Streamlit utility.
For deployments that require an API and real-time analysis, we offer an built-in mannequin internet hosting service, Metaflow Hosting. Although particulars have developed quite a bit, this outdated discuss nonetheless offers a very good overview of the service.
Metaflow Hosting is particularly geared in the direction of internet hosting artifacts or fashions produced in Metaflow. This gives a straightforward to make use of interface on high of Netflix’s present microservice infrastructure, permitting knowledge scientists to rapidly transfer their work from experimentation to a manufacturing grade internet service that may be consumed over a HTTP REST API with minimal overhead.
Its key advantages embody:
- Simple decorator syntax to create RESTFull endpoints.
- The back-end auto-scales the variety of cases used to again your service based mostly on visitors.
- The back-end will scale-to-zero if no requests are made to it after a specified period of time thereby saving value significantly in case your service requires GPUs to successfully produce a response.
- Request logging, alerts, monitoring and tracing hooks to Netflix infrastructure
Consider the service much like managed mannequin internet hosting providers like AWS Sagemaker Model Hosting, however tightly built-in with our microservice infrastructure.
Example use case: Media
We have an extended historical past of utilizing machine studying to course of media property, for example, to personalize art work and to assist our creatives create promotional content material effectively. Processing giant quantities of media property is technically non-trivial and computationally costly, so over time, we have now developed loads of specialised infrastructure devoted for this objective on the whole, and infrastructure supporting media ML use circumstances particularly.
To reveal the advantages of Metaflow Hosting that gives a general-purpose API layer supporting each synchronous and asynchronous queries, think about this use case involving Amber, our function retailer for media.
While Amber is a function retailer, precomputing and storing all media options upfront could be infeasible. Instead, we compute and cache options in an on-demand foundation, as depicted beneath:
When a service requests a function from Amber, it computes the function dependency graph after which sends a number of asynchronous requests to Metaflow Hosting, which locations the requests in a queue, finally triggering function computations when compute sources develop into obtainable. Metaflow Hosting caches the response, so Amber can fetch it after some time. We may have constructed a devoted microservice only for this use case, however because of the flexibleness of Metaflow Hosting, we had been capable of ship the function quicker with no further operational burden.
Our urge for food to use ML in numerous use circumstances is barely growing, so our Metaflow platform will maintain increasing its footprint correspondingly and proceed to offer pleasant integrations to programs constructed by different groups at Netlfix. For occasion, we have now plans to work on enhancements within the versioning layer, which wasn’t lined by this text, by giving extra choices for artifact and mannequin administration.
We additionally plan on constructing extra integrations with different programs which can be being developed by sister groups at Netflix. As an instance, Metaflow Hosting fashions are at present not nicely built-in into mannequin logging services — we plan on engaged on enhancing this to make fashions developed with Metaflow extra built-in with the suggestions loop important in coaching new fashions. We hope to do that in a pluggable method that will permit different customers to combine with their very own logging programs.
Additionally we need to provide extra methods Metaflow artifacts and fashions may be built-in into non-Metaflow environments and purposes, e.g. JVM based mostly edge service, in order that Python-based knowledge scientists can contribute to non-Python engineering programs simply. This would permit us to raised bridge the hole between the short iteration that Metaflow gives (in Python) with the necessities and constraints imposed by the infrastructure serving Netflix member dealing with requests.
If you might be constructing business-critical ML or AI programs in your group, be a part of the Metaflow Slack group! We are blissful to share experiences, reply any questions, and welcome you to contribute to Metaflow.
Acknowledgements:
Thanks to Wenbing Bai, Jan Florjanczyk, Michael Li, Aliki Mavromoustaki, and Sejal Rai for assist with use circumstances and figures. Thanks to our OSS contributors for making Metaflow a greater product.