Wednesday, February 17, 2016

Recommending for the World

#AlgorithmsEverywhere


Arabic.jpg


The Netflix experience is driven by a number of Machine Learning algorithms: personalized ranking, page generation, search, similarity, ratings, etc. On the 6th of January, we simultaneously launched Netflix in 130 new countries around the world, which brings the total to over 190 countries. Preparing for such a rapid expansion while ensuring each algorithm was ready to work seamlessly created new challenges for our recommendation and search teams. In this post, we highlight the four most interesting challenges we’ve encountered in making our algorithms operate globally and, most importantly, how this improved our ability to connect members worldwide with stories they'll love.


Challenge 1: Uneven Video Availability


Before we can add a video to our streaming catalog on Netflix, we need to obtain a license for it from the content owner. Most content licenses are region-specific or country-specific and are often held to terms for years at a time. Ultimately, our goal is to let members around the world enjoy all our content through global licensing, but currently our catalog varies between countries. For example, the dystopian Sci-Fi movie “Equilibrium” might be available on Netflix in the US but not in France. And “The Matrix” might be available in France but not in the US. Our recommendation models rely heavily on learning patterns from play data, particularly involving co-occurrence or sequences of plays between videos. In particular, many algorithms assume that when something was not played it is a (weak) signal that someone may not like a video, because they chose not to play it. However, in this particular scenario we will never observe any members who played both “Equilibrium” and “The Matrix”. A basic recommendation model would then learn that these two movies do not appeal to the same kinds of people just because the audiences were constrained to be different. However, if these two movies were available to the same set of members, we would likely observe a similarity between the videos and between the members who watch them. From this example, it is clear that uneven video availability potentially interferes with the quality of our recommendations.


Our search experience faces a similar challenge. Given a (partial) query from a member, we want to present the most relevant videos in the catalog. However, not accounting for availability differences reduces the quality of this ranking. For example, the top results for a given query from a ranking algorithm unaware of availability differences could include a niche video followed by a well-known one in a case where the latter is only available to a relatively small number of our global members and the former is available much more broadly.


Another aspect of content licenses is that they have start and end dates, which means that a similar problem arises not only across countries, but also within a given country across time. If we compare a well-known video that has only been available on Netflix for a single day to another niche video that was available for six months, we might conclude that the latter is a lot more engaging. However, if the recently added, well-known video had instead been on the site for six months, it probably would have more total engagement.


One can imagine the impact these issues can have on more sophisticated search or recommendation models when they already introduce a bias in something as simple as popularity. Addressing the issue of uneven availability across both geography and time lets our algorithms provide better recommendations for a video already on our service when it becomes available in a new country.


So how can we avoid learning catalog differences and focus on our real goal of learning great recommendations for our members? We incorporate into each algorithm the information that members have access to different catalogs based on geography and time, for example by building upon concepts from the statistical community on handling missing data.


Challenge 2: Cultural Awareness


Another key challenge in making our algorithms work well around the world is to ensure that we can capture local variations in taste. We know that even with the same catalog worldwide we would not expect a video to have the exact same popularity across countries. For example, we expect that Bollywood movies would have a different popularity in India than in Argentina. However, should two members get similar recommendations, if they have similar profiles but if one member lives in India and the other in Argentina? Perhaps if they are both watching a lot of Sci-Fi, their recommendations should be similar. Meanwhile, overall we would expect Argentine members should be recommended more Argentine Cinema and Indian members more Bollywood.


An obvious approach to capture local preferences would be to build models for individual countries. However, some countries are small and we will have very little member data available there. Training a recommendation algorithm on such sparse data leads to noisy results, as the model will struggle to identify clear personalization patterns from the data. So we need a better way.


Prior to our global expansion, our approach was to group countries into regions of a reasonable size that had a relatively consistent catalog and language. We would then build individual models for each region. This could capture the taste differences between regions because we trained separate models whose hyperparameters were tuned differently. Within a region, as long as there were enough members with certain taste preference and a reasonable amount of history, a recommendation model should be able to identify and use that pattern of taste. However, there were several problems with this approach. The first is that within a region the amount of data from a large country would dominate the model and dampen its ability to learn the local tastes for a country with a smaller number of members. It also presented a challenge of how to maintain the groupings as catalogs changed over time and memberships grew. Finally, because we’re continuously running A/B tests with model variants across many algorithms, the combinatorics involving a growing number of regions became overwhelming.


To address these challenges we sought to combine the regional models into a single global model that also improves the recommendations we make, especially in countries where we may not yet have many members. Of course, even though we are combining the data, we still need to reflect local differences in taste. This leads to the question: is local taste or personal taste more dominant? Based on the data we’ve seen so far, both aspects are important, but it is clear that taste patterns do travel globally. Intuitively, this makes sense: if a member likes Sci-Fi movies, someone on the other side of the world who also likes Sci-Fi would be a better source for recommendations than their next-door neighbor who likes food documentaries. Being able to discover worldwide communities of interest means that we can further improve our recommendations, especially for niche interests, as they will be based on more data. Then with a global algorithm we can identify new or different taste patterns that emerge over time.


To refine our models we can use many signals about the content and about our members. In this global context, two important taste signals could be language and location. We want to make our models aware of not just where someone is logged in from but also aspects of a video such as where it is from, what language it is in, and where it is popular. Going back to our example, this information would let us offer different recommendations to a brand new member in India as compared to Argentina, as the distribution of tastes within the two countries is different. We expand on the importance of language in the next section.


Challenge 3: Language


Netflix has now grown to support 21 languages and our catalog includes more local content than ever. This increase creates a number of challenges, especially for the instant search algorithm mentioned above. The key objective of this algorithm is to help every member find something to play whenever they search while minimizing the number of interactions. This is different than standard ranking metrics used to evaluate information retrieval systems, which do not take the amount of interaction into account. When looking at interactions, it is clear that different languages involve very different interaction patterns. For example, Korean is usually typed using the Hangul alphabet where syllables are composed from individual characters. For example, to search for “올드보이” (Oldboy), in the worst possible case, a member would have to enter nine characters: “ㅇ ㅗ ㄹㄷ ㅡ ㅂ ㅗ ㅇㅣ”. Using a basic indexing for the video title, in the best case a member would still need to type three characters: “ㅇ ㅗ ㄹ”, which would be collapsed in the first syllable of that title: “올”. In a Hangul-specific indexing, a member would need to write as little as one character: “ㅇ”. Optimizing for the best results with the minimum set of interactions and automatically adapting to newly introduced languages with significantly different writing systems is an area we’re working on improving.


Another language-related challenge relates to recommendations. As mentioned above, while taste patterns travel globally, ultimately people are most likely to enjoy content presented in a language they understand. For example, we may have a great French Sci-Fi movie on the service, but if there are no subtitles or audio available in English we wouldn’t want to recommend it to a member who likes Sci-Fi movies but only speaks English. Alternatively, if the member speaks both English and French, then there is a good chance it would be an appropriate recommendation. People also often have preferences for watching content that was originally produced in their native language, or one they are fluent in. While we constantly try to add new language subtitles and dubs to our content, we do not yet have all languages available for all content. Furthermore, different people and cultures also have different preferences for watching with subtitles or dubs. Putting this together, it seems clear that recommendations could be better with an awareness of language preferences. However, currently which languages a member understands and to what degree is not defined explicitly, so we need to infer it from ancillary data and viewing patterns.


Challenge 4: Tracking Quality


The objective is to build recommendation algorithms that work equally well for all of our members; no matter where they live or what language they speak. But with so many members in so many countries speaking so many languages, a challenge we now face is how to even figure out when an algorithm is sub-optimal for some subset of our members.


To handle this, we could use some of the approaches for the challenges above. For example, we could look at the performance of our algorithms by manually slicing along a set of dimensions (country, language, catalog, …). However, some of these slices lead to very sparse and noisy data. At the other end of the scale we could be looking at metrics observed globally, but this would dramatically limit our ability to detect issues until they impact a large number of our members. One approach this problem is to learn how to best group observations for the purpose of automatically detecting outliers and anomalies. Just as we work on improving our recommendation algorithms, we are innovating our metrics, instrumentation and monitoring to improve their fidelity and through them our ability to detect new problems and highlight areas to improve our service.


Conclusion

To support a launch of this magnitude, we examined each and every algorithm that is part of our service and began to address these challenges. Along the way, we found not just approaches that will make Netflix better for those signing up in the 130 new countries, but in fact better for all Netflix members worldwide. For example, solving the first and the second challenges let us discover worldwide communities of interest so that we can make better recommendations. Solving the third challenge means that regardless of where our members are based, they can use Netflix in the language that suits them the best, and quickly find the content they’re looking for. Solving the fourth challenge means that we’re able to detect issues at a finer grain and so that our recommendation and search algorithms help all our members find content they love. Of course, our global journey is just beginning and we look forward to making our service dramatically better over time. If you are an algorithmic explorer who finds this type of adventure exciting, take a look at our current job openings.

Monday, February 15, 2016

Evolution of the Netflix Data Pipeline

Our new Keystone data pipeline went live in December of 2015. In this article, we talk about the evolution of Netflix’s data pipeline over the years. This is the first of a series of articles about the new Keystone data pipeline.

Netflix is a data-driven company. Many business and product decisions are based on insights derived from data analysis. The charter of the data pipeline is to collect, aggregate, process and move data at cloud scale. Almost every application at Netflix uses the data pipeline.

Here are some statistics about our data pipeline:
  • ~500 billion events and  ~1.3 PB per day 
  • ~8 million events and ~24 GB per second during peak hours

There are several hundred event streams flowing through the pipeline. For example:
  • Video viewing activities
  • UI activities
  • Error logs
  • Performance events
  • Troubleshooting & diagnostic events

Note that operational metrics don’t flow through this data pipeline. We have a separate telemetry system Atlas, which we open-sourced just like many other Netflix technologies.

Over the last a few years, our data pipeline has experienced major transformations due to evolving requirements and technological developments.

V1.0 Chukwa pipeline

The sole purpose of the original data pipeline was to aggregate and upload events to Hadoop/Hive for batch processing. As you can see, the architecture is rather simple. Chukwa collects events and writes them to S3 in Hadoop sequence file format. The Big Data Platform team further processes those S3 files and writes to Hive in Parquet format. End-to-end latency is up to 10 minutes. That is sufficient for batch jobs which usually scan data at daily or hourly frequency.
V1.0- Chukwa Pipeline (1).png

V1.5 Chukwa pipeline with real-time branch

With the emergence of Kafka and Elasticsearch over the last couple of years, there has been a growing demand for real-time analytics in Netflix. By real-time, we mean sub-minute latency.
V1.5- Chukwa Pipeline with Real-Time Branch.jpg
In addition to uploading events to S3/EMR, Chukwa can also tee traffic to Kafka (the front gate of real-time branch). In V1.5, approximately 30% of the events are branched to the real-time pipeline. The centerpiece of the real-time branch is the router. It is responsible for routing data from Kafka to the various sinks: Elasticsearch or secondary Kafka.

We have seen explosive growth in Elasticsearch adoption within Netflix for the last two years. There are ~150 clusters totaling ~3,500 instances hosting ~1.3 PB of data. The vast majority of the data is injected via our data pipeline.  

When Chukwa tees traffic to Kafka, it can deliver full or filtered streams. Sometimes, we need to apply further filtering on the Kafka streams written from Chukwa. That is why we have the router to consume from one Kafka topic and produce to a different Kafka topic.

Once we deliver data to Kafka, it empowers users with real-time stream processing: Mantis, Spark, or custom applications. “Freedom and Responsibility” is the DNA of Netflix culture. It’s up to users to choose the right tool for the task at hand.

Because moving data at scale is our expertise, our team maintains the router as a managed service. But there are a few lessons we learned while operating the routing service:
  • The Kafka high-level consumer can lose partition ownership and stop consuming some partitions after running stable for a while. This requires us to bounce the processes.
  • When we push out new code, sometimes the high-level consumer can get stuck in a bad state during rebalance.
  • We group hundreds of routing jobs into a dozen of clusters. The operational overhead of managing those jobs and clusters is an increasing burden. We need a better platform to manage the routing jobs.

V2.0 Keystone pipeline (Kafka fronted)

In addition to the issues related to routing service, there are other motivations for us to revamp our data pipeline:
  • Simplify the architecture.
  • Kafka implements replication that improves durability, while Chukwa doesn’t support replication.
  • Kafka has a vibrant community with strong momentum.

V2.0 Keystone Pipeline (1).png

There are three major components:

  • Data Ingestion - There are two ways for applications to ingest data.
    • use our Java library and write to Kafka directly.
    • send to an HTTP proxy which then writes to Kafka.
  • Data Buffering - Kafka serves as the replicated persistent message queue. It also helps absorb temporary outages from downstream sinks.
  • Data Routing - The routing service is responsible for moving data from fronting Kafka to various sinks: S3, Elasticsearch, and secondary Kafka.

We have been running Keystone pipeline in production for the past few months. We are still evolving Keystone with a focus on QoS, scalability, availability, operability, and self-service.

In follow-up posts, we’ll cover more details regarding:
  • How do we run Kafka in cloud at scale?
  • How do we implement routing service using Samza?
  • How do we manage and deploy Docker containers for routing service?

If building large-scale infrastructure excites you, we are hiring!

Real-Time Data Infrastructure Team

Friday, February 12, 2016

Distributed Time Travel for Feature Generation


We want to make it easy for Netflix members to find great content to fulfill their unique tastes. To do this, we follow a data-driven algorithmic approach based on machine learning, which we have described in past posts and other publications. We aspire to a day when anyone can sit down, turn on Netflix, and the absolute best content for them will automatically start playing. While we are still way off from that goal, it sets a vision for us to improve the algorithms that span our service: from how we rank videos to how we construct the homepage to how we provide search results. To make our algorithms better, we follow a two-step approach. First, we try an idea offline using historical data to see if it would have made better recommendations. If it does, we then deploy a live A/B test to see if it performs well in reality, which we measure through statistically significant improvements in core metrics such as member engagement, satisfaction, and retention.

While there are many ways to improve machine learning approaches, arguably the most critical is to provide better input data. A model can only be as good as the data we give it. Thus, we spend a lot of time experimenting with new kinds of input signals for our models. Most machine learning models expect input to be represented as a vector of numbers, known as a feature vector. Somehow we need to take an arbitrary input entity (e.g. a tuple of member profile, video, country, time, device, etc.), with its associated, richly structured data, and provide a feature vector representing that entity for a machine learning algorithm to use. We call this transformation feature generation and it is central to providing the data needed for learning. Examples features include how many minutes a member has watched a video, the popularity of the video, its predicted rating, what genre a video belongs to, or how many videos are in a row. We use the term feature broadly, since a feature could be a simple indicator or have a full model behind it, such as a Matrix Factorization.

We will describe how we built a time machine for feature generation using Apache Spark that enables our researchers to quickly try ideas for new features on historical data such that running offline experiments and transitioning to online A/B tests is seamless.

Why build a time machine?

There are many ways to approach feature generation, several of which we’ve used in the past. One way is to use logged event data that we store on S3 and access via Hive by running queries on these tables to define features. While this is flexible for exploratory analysis, it has several problems. First, to run an A/B test we need the feature calculation to run within our online microservice architecture. We run the models online because we know that freshness and responsiveness of our recommendations is important to the member experience. This means we would need to re-implement feature generation to retrieve data from online services instead of Hive tables. It is difficult to match two such implementations exactly, especially since any discrepancies between offline and online data sources can create unexpected differences in the model output. In addition, not all of our data is available offline, particularly output of recommendation models, because these involve a sparse-to-dense conversion that creates a large volume of data.

On the other extreme, we could log our features online where a model would be used. While this removes the offline/online discrepancy and makes transitioning to A/B test easy, it means we need to deploy each idea for a new feature into production and wait for the data to collect before we can determine if a feature is useful. This slows down the iteration cycle for new ideas. It also requires that all the data for a feature to be available online, which could mean building new systems to serve that data, again before we have determined if it is valuable. We also need to compute features for many more members or requests than we may actually need for training based on how we choose label data.

We’ve also tried a middle ground where we use feature code that calls online services, such as the one that provides viewing history, and filters out all the data with timestamps past a certain point in time. However, this only works for situations where a service records a log of all historical events; services that just provide the current state cannot be used. It also places additional load on the online services each time we generate features.

Throughout these approaches, management of time is extremely important. We want an approach that balances the benefits of all the above approaches without the drawbacks. In particular, we want a system that:
  • Enables quick iteration from idea to modeling to running an A/B test
  • Uses the data provided by our online microservices, without overloading them
  • Accurately represents input data for a model at a point in time to simulate online use
  • Handles our scale of data with many researchers running experiments concurrently, without using more than 1.21 gigawatts of power
  • Works well in an interactive environment, such as using a notebook for experimentation, and also reliably in a batch environment, such as for doing periodic retraining
  • Should only need to write feature code once so that we don’t need to spend time verifying that two implementations are exactly equivalent
  • Most importantly, no paradoxes are allowed (e.g. the label can’t be in the features)
When faced with tough problems one often wishes for a time machine to solve them. So that is what we decided to build. Our time machine snapshots online services and uses the snapshot data offline to reconstruct the inputs that a model would have seen online to generate features. Thus, when experimenters design new feature encoders — functions that take raw data as input and compute features — they can immediately use them to compute new features for any time in the past, since the time machine can retrieve the appropriate snapshots and pass them to the feature encoders.

How to build a Time Machine

Here are the various components needed in a time machine that snapshots online services:
  • Select contexts to snapshot
  • Snapshot data of various micro services for the selected context
  • Build APIs to serve this data for a given time coordinate in the past

Context Selection

Snapshotting data for all contexts (e.g all member profiles, devices, times of day) would be very expensive. Instead, we select samples of contexts to snapshot periodically (typically daily), though different algorithms may need to train on different distributions. For example, some use  stratified samples based on properties such as viewing patterns, devices, time spent on the service, region, etc. To handle this, we use Spark SQL to select an appropriate sample of contexts for each experiment from Hive. We merge the context set across experiments and persist it into S3 along with the corresponding experiment identifiers.

Data Snapshots

The next component in the time machine fetches data from various online services and saves a snapshot of the returned data for the selected contexts. Netflix embraces a fine-grained Service Oriented Architecture for our cloud-based deployment model. There are hundreds of such micro services that are collectively responsible for handling the member experience. Data from various such services like Viewing History, My List, and Predicted Ratings are used as input for the features in our models.

We use Netflix-specific components such as Eureka, Hystrix, and Archaius to fetch data from online services through their client libraries. However, some of these client libraries bulk-load data, so they have a high memory footprint and a large startup time. Spark is not well suited for loading such components inside its JVM. Moreover, the requirement of creating an uber jar to run Spark jobs can cause runtime jar incompatibility issues with other Netflix libraries. To alleviate this problem, we used Prana, which runs outside the Spark JVM, as a data proxy to the Netflix ecosystem.

Spark parallelizes the calls to Prana, which internally fetches data from various micro services for each of these contexts. We chose Thrift as the binary communication protocol between Spark and Prana. We store the snapshotted data in S3 using Parquet, a compressed column-oriented binary format, for both time and space efficiency, and persist the location of the S3 data in Cassandra.
Ensuring pristine data quality of these snapshots is critical for us to correctly evaluate our models. Hence, we store the confidence level for each snapshot service, which is the percentage of successful data fetches from the micro services excluding any fallbacks due to timeouts or service failures. We expose it to our clients, who can chose to use this information for their experimentation.

For both snapshotting and context selection, we needed to schedule several Spark jobs to run on a periodic basis, with dependencies between them. To that end, we built a general purpose workflow orchestration and scheduling framework called Meson, which is optimized for machine learning pipelines, and used it to run the Spark jobs for the components of the time machine. We intend to open source Meson in the future and will provide more detail about it in an upcoming blog post.

APIs for Time Travel

We built APIs that enable time travel and fetch the snapshot data from S3 for a given time in the past. Here is a sample API to get the snapshot data for the Viewing History service.






Given a destination time in the past, the API fetches the associated S3 location of the snapshot data from Cassandra and loads the snapshot data in Spark. In addition, when given an A/B test identifier, the API filters the snapshot data to return only those contexts selected for that A/B test. The system transforms the snapshot data back into the respective services’ Java objects (POJOs) so that the feature encoders operate on the exact same POJOs for both offline experimentation and online feature generation in production.

The following diagram shows the overall architecture of the time machine and where Spark is used in building it: from selecting members for experimentation, snapshotting data of various services for the selected members, to finally serving the data for a time in the past.


DeLorean: Generating Features via Time Travel

DeLorean is our internal project to build the system that takes an experiment plan, travels back in time to collect all the necessary data from the snapshots, and generates a dataset of features and labels for that time in the past to train machine learning models. Of course, the first step is to select the destination time, to bring it up to 88 miles per hour, then DeLorean takes care of the rest.

Running an Experiment

DeLorean allows a researcher to run a wide range of experiments by automatically determining how to launch the time machine, what time coordinates are needed, what data to retrieve, and how to structure the output. Thus, to run a new experiment, an experimenter only needs to provide the following:
  • Label data: A blueprint for obtaining a set of contexts with associated time coordinates, items, and labels for each. This is typically created by a Hive, Pig, or Spark SQL query
  • A feature model containing the required feature encoder configurations
  • Implementations of any new feature encoders that do not already exist in our library
DeLorean provides a capability for writing and modifying a new feature encoder during an experiment, for example, in a Zeppelin Notebook or in Spark Shell, so that it can be used immediately for feature generation. If we find that new feature encoder useful, we can later productionize it by adding it to our library of feature encoders.

The high-level process to generate features is depicted in the following diagram, where the blocks highlighted in light green are typically customized for new experiments. In this scenario, experimenters can also implement new feature encoders that are used in conjunction with existing ones.
DeLorean image by JMortonPhoto.com & OtoGodfrey.com

Label Data and Feature Encoders

One of the primary inputs to DeLorean is the label data, which contains information about the contexts, items, and associated labels for which to generate features. The contexts, as its name suggests, can be describe the setting for where a model is to be used (e.g. tuples of member profiles, country, time, device, etc.). Items are the elements which are to be trained on, scored, and/or ranked (e.g. videos, rows, search entities). Labels are typically the targets used in supervised learning for each context-item combination. For unsupervised learning approaches, the label is not required. As an example, for personalized ranking the context could be defined as the member profile ID, country code, and time, whereas the item as the video, and the labels as plays or non-plays. In this example, the label data is created by joining the set of snapshotted contexts to the logged play actions.

Once we have this label dataset, we need to compute features for each context-item combination in the dataset by using the desired set of feature encoders. Each feature encoder takes a context and each of target items associated with the context, together with some raw data elements in the form of POJOs, to compute one or more features.

Each type of item, context variable or data element, has a data key associated with it. Every feature encoder has a method that returns the set of keys for the data it consumes. DeLorean uses these keys to identify the required data types, retrieves the data, and passes it to the feature encoder as a data map — which is a map from data keys to data objects.

We made DeLorean flexible enough to allow the experiments to use different types of contexts and items without needing to customize the feature generation system. DeLorean can be used not only for recommendations, but also for a row ordering experiment which has profile-device tuple as context and rows of videos as items. Another use case may be a search experiment which has the query-profile-country tuple as context and individual videos as items. To achieve this, DeLorean automatically infers the type of contexts and items from the label data and the data keys required by the feature encoders.

Data Elements

Data elements are the ingredients that get transformed into features by a feature encoder. Some of these are context-dependent, such as viewing history for a profile, and others are shared by all contexts, such as metadata of the videos. We handle these two types of data elements differently.

For context-dependent data elements, we use the snapshots described above, and associate each one with a data key. We bring all the required snapshot data sources together with the values, items, and labels for each context, so that the data for a single context is sent to a single Spark executor. Different contexts are broken up to enable distributed feature generation. The snapshots are loaded as an RDD of (context, Map(data key -> data element)) in a lazy fashion and a series of joins between the label data and all the necessary context-dependent data elements are performed using Spark.

For context-independent data elements, DeLorean broadcasts these bulk data elements to each executor. Since these data elements have manageable sizes and often have a slow rate of change over time, we keep a record of each update that we use to rewind back to the appropriate previous version. These are kept in memory as singleton objects and made available to the feature generators for each context processed by an executor. Thus, a complete data map is created for each context containing the context data, context-dependent snapshot data elements, and shared data singletons.

Once the features are generated in Spark, the data is represented as a Spark DataFrame with an embedded schema. For many personalization application, we need to rank a number of items for each context. To avoid shuffling in the ranking process, item features are grouped by context in the output. The final features are stored in Hive using a Parquet format.

Model Training, Validation, and Testing

We use features generated using our time machine to train the models that we use in various parts of our recommendation systems. We use a standardized schema for passing the DataFrames of training features to machine learning algorithms, as well as computing predictions and metrics for trained models on the validation and test feature DataFrames. We also standardized a format to serialize the models that we use for publishing the models to be later consumed by online applications or in other future experiments.

The following diagram shows how we run a typical machine learning experiment. Once the experiment is designed, we collect the dataset of contexts, items, and labels. Next the features for the label dataset are generated. We then train models using either single machine, multi-core, or distributed algorithms and perform parameter tuning by computing metrics on a validation set. Then we pick the best models and compare them on a testing set. When we see a significant improvement in the offline metrics over the production model and that the outputs are different enough, we design an A/B test using variations of the model and run it online. If the A/B test shows a statistically significant increase in core metrics, we roll it out broadly. Otherwise, we learn from the results to iterate on the next idea.  

Going Online

One of the primary motivations for building DeLorean is to share the same feature encoders between offline experiments and online scoring systems to ensure that there are no discrepancies between the features generated for training and those computed online in production. When an idea is ready to be tested online, the model is packaged with the same feature configuration that was used by DeLorean to generate the features.

To compute features in the production system, we directly call our online microservices to collect the data elements required by all the feature encoders used in a model, instead of obtaining them from snapshots as we do offline. We then assemble them into data maps and pass them to the feature encoders. The feature vector is then passed to the offline-trained model for computing predictions, which are used to create our recommendations. The following diagram shows the high-level process of transitioning from an offline experiment to an online production system where the blocks highlighted in yellow are online systems, and the ones highlighted in blue are offline systems. Note that the feature encoders are shared between online and offline to guarantee the consistency of feature generation.

Conclusion and Future work

By collecting the state of the online world at a point in time for a select set of contexts, we were able to build a mechanism for turning back time. Spark’s distributed, resilient computation power enabled us to snapshot millions of contexts per day and to implement feature generation, model training and validation at scale. DeLorean is now being used in production for feature generation in some of the latest A/B tests for our recommender system.

However, this is just a start and there are many ways in which we can improve this approach. Instead of batch snapshotting on a periodic cadence, we can drive the snapshots based on events, for example at a time when a particular member visits our service. To avoid duplicate data collection, we can also capture data changes instead of taking full snapshots each time. We also plan on using the time machine capability for other needs in evaluating new algorithms and testing our systems. Of course, we leave the ability to travel forward in time as future work.

Fast experimentation is the hallmark of a culture of innovation. Reducing the time to production for an idea is a key metric we use to measure the success of our infrastructure projects. We will continue to build on this foundation to bring better personalization to Netflix in our effort to delight members and win moments of truth. If you are interested in these types of time-bending engineering challenges, join us.