April 23, 2024

At Netflix, to advertise and suggest the content material to customers in the very best approach there are a lot of Media Algorithm groups which work hand in hand with content material creators and editors. A number of of those algorithms goal to enhance totally different handbook workflows in order that we present the customized promotional picture, trailer or the present to the consumer.

These media centered machine studying algorithms in addition to different groups generate a variety of information from the media recordsdata, which we described in our earlier weblog, are saved as annotations in Marken. We designed a singular idea referred to as Annotation Operations which permits groups to create information pipelines and simply write annotations with out worrying about entry patterns of their information from totally different purposes.

Annotation Operations

Lets decide an instance use case of figuring out objects (like bushes, vehicles and many others.) in a video file. As described within the above image

  • Through the first run of the algorithm it recognized 500 objects in a selected Video file. These 500 objects have been saved as annotations of a particular schema sort, let’s say Objects, in Marken.
  • The Algorithm workforce improved their algorithm. Now once we re-ran the algorithm on the identical video file it created 600 annotations of schema sort Objects and saved them in our service.

Discover that we can not replace the annotations from earlier runs as a result of we don’t know what number of annotations a brand new algorithm run will outcome into. Additionally it is very costly for us to maintain observe of which annotation must be up to date.

The objective is that when the patron comes and searches for annotations of sort Objects for the given video file then the next ought to occur.

  • Earlier than Algo run 1, in the event that they search they need to not discover something.
  • After the completion of Algo run 1, the question ought to discover the primary set of 500 annotations.
  • Through the time when Algo run 2 was creating the set of 600 annotations, shoppers search ought to nonetheless return the older 500 annotations.
  • When the entire 600 annotations are efficiently created, they need to change the older set of 500.
  • So now when shoppers search annotations for Objects then they need to get 600 annotations.

Does this remind you of one thing? This appears very related (not precisely identical) to a distributed transaction.

Usually, an algorithm run can have 2k-5k annotations. There are various naive options doable for this drawback for instance:

  • Write totally different runs in several databases. That is clearly very costly.
  • Write algo runs into recordsdata. However we can not search or current low latency retrievals from recordsdata
  • And so on.

As a substitute our problem was to implement this characteristic on high of Cassandra and ElasticSearch databases as a result of that’s what Marken makes use of. The answer which we current on this weblog shouldn’t be restricted to annotations and can be utilized for every other area which makes use of ES and Cassandra as effectively.

Marken’s structure diagram is as follows. We refer the reader to our earlier weblog article for particulars. We use Cassandra as a supply of reality the place we retailer the annotations whereas we index annotations in ElasticSearch to offer wealthy search functionalities.

Marken Structure

Our objective was to assist groups at Netflix to create information pipelines with out serious about how that information is offered to the readers or the consumer groups. Equally, consumer groups don’t have to fret about when or how the information is written. That is what we name decoupling producer flows from shoppers of the information.

Lifecycle of a film goes by a variety of artistic levels. We’ve many non permanent recordsdata that are delivered earlier than we get to the ultimate file of the film. Equally, a film has many alternative languages and every of these languages can have totally different recordsdata delivered. Groups usually wish to run algorithms and create annotations utilizing all these media recordsdata.

Since algorithms will be run on a special permutations of how the media recordsdata are created and delivered we are able to simplify an algorithm run as follows

  • Annotation Schema Sort — identifies the schema for the annotation generated by the Algorithm.
  • Annotation Schema Model — identifies the schema model of the annotation generated by the Algorithm.
  • PivotId — a singular string identifier which identifies the file or technique which is used to generate the annotations. This could possibly be the SHA hash of the file or just the film Identifier quantity.

Given above we are able to describe the information mannequin for an annotation operation as follows.

"annotationOperationKeys": [

"annotationType": "string", ❶
"annotationTypeVersion": “integer”,
"pivotId": "string",
"operationNumber": “integer” ❷

"id": "UUID",
"operationStatus": "STARTED", ❸
"isActive": true ❹
  1. We already defined AnnotationType, AnnotationTypeVersion and PivotId above.
  2. OperationNumber is an auto incremented quantity for every new operation.
  3. OperationStatus — An operation goes by three phases, Began, Completed and Canceled.
  4. IsActive — Whether or not an operation and its related annotations are energetic and searchable.

As you may see from the information mannequin that the producer of an annotation has to decide on an AnnotationOperationKey which lets them outline how they need UPSERT annotations in an AnnotationOperation. Inside, AnnotationOperationKey the vital discipline is pivotId and the way it’s generated.

Our supply of reality for all objects in Marken in Cassandra. To retailer Annotation Operations we now have the next principal tables.

  • AnnotationOperationById — It shops the AnnotationOperations
  • AnnotationIdByAnnotationOperationId — it shops the Ids of all annotations in an operation.

Since Cassandra is NoSql, we now have extra tables which assist us create reverse indices and run admin jobs in order that we are able to scan all annotation operations every time there’s a want.

Every annotation in Marken can also be listed in ElasticSearch for powering varied searches. To file the connection between annotation and operation we additionally index two fields

  • annotationOperationId — The ID of the operation to which this annotation belongs
  • isAnnotationOperationActive — Whether or not the operation is in an ACTIVE state.

We offer three APIs to our customers. In following sections we describe the APIs and the state administration finished throughout the APIs.


When this API known as we retailer the operation with its OperationKey (tuple of annotationType, annotationType Model and pivotId) in our database. This new operation is marked to be in STARTED state. We retailer all OperationIDs that are in STARTED state in a distributed cache (EVCache) for quick entry throughout searches.



Customers name this API to upsert the annotations in an Operation. They move annotations together with the OperationID. We retailer the annotations and in addition file the connection between the annotation IDs and the Operation ID in Cassandra. Throughout this section operations are in isAnnotationOperationActive = ACTIVE and operationStatus = STARTED state.

Be aware that sometimes in a single operation run there will be 2K to 5k annotations which will be created. Shoppers can name this API from many alternative machines or threads for quick upserts.



As soon as the annotations have been created in an operation shoppers name FinishAnnotationOperation which adjustments following

  • Marks the present operation (let’s say with ID2) to be operationStatus = FINISHED and isAnnotationOperationActive=ACTIVE.
  • We take away the ID2 from the Memcache since it’s not in STARTED state.
  • Any earlier operation (let’s say with ID1) which was ACTIVE is now marked isAnnotationOperationActive=FALSE in Cassandra.
  • Lastly, we name updateByQuery API in ElasticSearch. This API finds all Elasticsearch paperwork with ID1 and marks isAnnotationOperationActive=FALSE.

Search API

That is the important thing half for our readers. When a consumer calls our search API we should exclude

  • any annotations that are from isAnnotationOperationActive=FALSE operations or
  • for which Annotation operations are at present in STARTED state. We try this by excluding the next from all queries in our system.

To realize above

  1. We add a filter in our ES question to exclude isAnnotationOperationStatus is FALSE.
  2. We question EVCache to search out out all operations that are in STARTED state. Then we exclude all these annotations with annotationId present in memcache. Utilizing memcache permits us to maintain latencies for our search low (most of our queries are lower than 100ms).

Cassandra is our supply of reality so if an error occurs we fail the consumer name. Nevertheless, as soon as we decide to Cassandra we should deal with Elasticsearch errors. In our expertise, all errors have occurred when the Elasticsearch database is having some concern. Within the above case, we created a retry logic for updateByQuery calls to ElasticSearch. If the decision fails we push a message to SQS so we are able to retry in an automatic vogue after some interval.

In close to time period, we wish to write a excessive degree abstraction single API which will be referred to as by our shoppers as a substitute of calling three APIs. For instance, they will retailer the annotations in a blob storage like S3 and provides us a hyperlink to the file as a part of the only API.