Scalable Annotation Service — Marken | Netflix TechBlog

At Netflix, now we have a whole bunch of micro companies every with its personal knowledge fashions or entities. For instance, now we have a service that shops a film entity’s metadata or a service that shops metadata about photos. All of those companies at a later level wish to annotate their objects or entities. Our crew, Asset Administration Platform, determined to create a generic service referred to as Marken which permits any microservice at Netflix to annotate their entity.


Typically folks describe annotations as tags however that may be a restricted definition. In Marken, an annotation is a bit of metadata which may be hooked up to an object from any area. There are a lot of completely different sorts of annotations our consumer functions wish to generate. A easy annotation, like under, would describe {that a} explicit film has violence.

  • Film Entity with id 1234 has violence.

However there are extra attention-grabbing circumstances the place customers wish to retailer temporal (time-based) knowledge or spatial knowledge. In Pic 1 under, now we have an instance of an utility which is utilized by editors to evaluation their work. They wish to change the colour of gloves to wealthy black so they need to have the ability to mark up that space, on this case utilizing a blue circle, and retailer a remark for it. This can be a typical use case for a inventive evaluation utility.

An instance for storing each time and area based mostly knowledge can be an ML algorithm that may determine characters in a body and needs to retailer the next for a video

  • In a specific body (time)
  • In some space in picture (area)
  • A personality title (annotation knowledge)
Pic 1 : Editors requesting modifications by drawing shapes just like the blue circle proven above.

Objectives for Marken

We wished to create an annotation service which can have the next targets.

  • Permits to annotate any entity. Groups ought to be capable to outline their knowledge mannequin for annotation.
  • Annotations may be versioned.
  • The service ought to be capable to serve real-time, aka UI, functions so CRUD and search operations must be achieved with low latency.
  • All knowledge must be additionally obtainable for offline analytics in Hive/Iceberg.


Because the annotation service can be utilized by anybody at Netflix we had a have to help completely different knowledge fashions for the annotation object. An information mannequin in Marken may be described utilizing schema — similar to how we create schemas for database tables and many others.

Our crew, Asset Administration Platform, owns a distinct service that has a json based mostly DSL to explain the schema of a media asset. We prolonged this service to additionally describe the schema of an annotation object.

"sort": "BOUNDING_BOX", ❶
"model": 0, ❷
"description": "Schema describing a bounding field",
"properties": ❸
"sort": "bounding_box",
"necessary": true
"sort": "time_range",
"necessary": true

Within the above instance, the appliance desires to characterize in a video an oblong space which spans a variety of time.

  1. Schema’s title is BOUNDING_BOX
  2. Schemas can have variations. This permits customers to make add/take away properties of their knowledge mannequin. We don’t permit incompatible modifications, for instance, customers can’t change the information sort of a property.
  3. The information saved is represented within the “properties” part. On this case, there are two properties
  4. boundingBox, with sort “bounding_box”. That is principally an oblong space.
  5. boxTimeRange, with sort “time_range”. This permits us to specify begin and finish time for this annotation.

Geometry Objects

To characterize spatial knowledge in an annotation we used the Well Known Text (WKT) format. We help following objects

  • Level
  • Line
  • MultiLine
  • BoundingBox
  • LinearRing

Our mannequin is extensible permitting us to simply add extra geometry objects as wanted.

Temporal Objects

A number of functions have a requirement to retailer annotations for movies which have time in it. We permit functions to retailer time as body numbers or nanoseconds.

To retailer knowledge in frames purchasers should additionally retailer frames per second. We name this a SampleData with following parts:

  • sampleNumber aka body quantity
  • sampleNumerator
  • sampleDenominator

Annotation Object

Similar to schema, an annotation object can be represented in JSON. Right here is an instance of annotation for BOUNDING_BOX which we mentioned above.

"annotationId": ❶
"id": "188c5b05-e648-4707-bf85-dada805b8f87",
"model": "0"
"associatedId": ❷
"entityType": "MOVIE_ID",
"id": "1234"
"annotationType": "ANNOTATION_BOUNDINGBOX", ❸
"annotationTypeVersion": 1,
"metadata": ❹
"fileId": "identityOfSomeFile",
"x": 20,
"y": 30
"x": 40,
"y": 60

"startTimeInNanoSec": 566280000000,
"endTimeInNanoSec": 567680000000

  1. The primary element is the distinctive id of this annotation. An annotation is an immutable object so the id of the annotation all the time features a model. At any time when somebody updates this annotation we mechanically increment its model.
  2. An annotation have to be related to some entity which belongs to some microservice. On this case, this annotation was created for a film with id “1234”
  3. We then specify the schema sort of the annotation. On this case it’s BOUNDING_BOX.
  4. Precise knowledge is saved within the metadata part of json. Like we mentioned above there’s a bounding field and time vary in nanoseconds.

Base schemas

Similar to in Object Oriented Programming, our schema service permits schemas to be inherited from one another. This permits our purchasers to create an “is-a-type-of” relationship between schemas. Not like Java, we help a number of inheritance as effectively.

We have now a number of ML algorithms which scan Netflix media property (photos and movies) and create very attention-grabbing knowledge for instance figuring out characters in frames or figuring out match cuts. This knowledge is then saved as annotations in our service.

As a platform service we created a set of base schemas to ease creating schemas for various ML algorithms. One base schema (TEMPORAL_SPATIAL_BASE) has the next elective properties. This base schema can be utilized by any derived schema and never restricted to ML algorithms.

  • Temporal (time associated knowledge)
  • Spatial (geometry knowledge)

And one other one BASE_ALGORITHM_ANNOTATION which has the next elective properties which is often utilized by ML algorithms.

  • label (String)
  • confidenceScore (double) — denotes the arrogance of the generated knowledge from the algorithm.
  • algorithmVersion (String) — model of the ML algorithm.

By utilizing a number of inheritance, a typical ML algorithm schema derives from each TEMPORAL_SPATIAL_BASE and BASE_ALGORITHM_ANNOTATION schemas.

"model": 0,
"description": "Base Schema for Algorithm based mostly Annotations",
"sort": "decimal",
"necessary": false,
"description": "Confidence Rating",
"sort": "string",
"necessary": false,
"description": "Annotation Tag",
"sort": "string",
"description": "Algorithm Model"


Given the targets of the service we needed to maintain following in thoughts.

  • Our service shall be utilized by lots of inside UI functions therefore the latency for CRUD and search operations have to be low.
  • Moreover functions we can have ML algorithm knowledge saved. A few of this knowledge may be on the body degree for movies. So the quantity of knowledge saved may be giant. The databases we decide ought to be capable to scale horizontally.
  • We additionally anticipated that the service can have excessive RPS.

Another targets got here from search necessities.

  • Capability to go looking the temporal and spatial knowledge.
  • Capability to go looking with completely different related and extra related Ids as described in our Annotation Object knowledge mannequin.
  • Full textual content searches on many alternative fields within the Annotation Object
  • Stem search help

As time progressed the necessities for search solely elevated and we are going to talk about these necessities intimately in a distinct part.

Given the necessities and the experience in our crew we determined to decide on Cassandra because the supply of fact for storing annotations. For supporting completely different search necessities we selected ElasticSearch. Moreover to help numerous options now we have bunch of inside auxiliary companies for eg. zookeeper service, internationalization service and many others.

Marken structure

Above image represents the block diagram of the structure for our service. On the left we present knowledge pipelines that are created by a number of of our consumer groups to mechanically ingest new knowledge into our service. A very powerful of such a knowledge pipeline is created by the Machine Studying crew.

One of many key initiatives at Netflix, Media Search Platform, now makes use of Marken to retailer annotations and carry out numerous searches defined under. Our structure makes it doable to simply onboard and ingest knowledge from Media algorithms. This knowledge is utilized by numerous groups for eg. creators of promotional media (aka trailers, banner photos) to enhance their workflows.


Success of Annotation Service (knowledge labels) will depend on the efficient search of these labels with out realizing a lot of enter algorithms particulars. As talked about above, we use the bottom schemas for each new annotation sort (relying on the algorithm) listed into the service. This helps our purchasers to go looking throughout the completely different annotation varieties persistently. Annotations may be searched both by merely knowledge labels or with extra added filters like film id.

We have now outlined a customized question DSL to help looking, sorting and grouping of the annotation outcomes. Several types of search queries are supported utilizing the Elasticsearch as a backend search engine.

  • Full Textual content Search — Purchasers could not know the precise labels created by the ML algorithms. For instance, the label may be ‘bathe curtain’. With full textual content search, purchasers can discover the annotation by looking utilizing label ‘curtain’ . We additionally help fuzzy search on the label values. For instance, if the purchasers wish to search ‘curtain’ however they wrongly typed ‘curtian` — annotation with the ‘curtain’ label shall be returned.
  • Stem Search — With world Netflix content material supported in several languages, our purchasers have the requirement to help stem seek for completely different languages. Marken service incorporates subtitles for a full catalog of titles in Netflix which may be in many alternative languages. For instance for stem search , `clothes` and `garments` may be stemmed to the identical root phrase `fabric`. We use ElasticSearch to help stem seek for 34 completely different languages.
  • Temporal Annotations Search — Annotations for movies are extra related whether it is outlined together with the temporal (time vary with begin and finish time) data. Time vary inside video can be mapped to the body numbers. We help labels seek for the temporal annotations inside the supplied time vary/body quantity additionally.
  • Spatial Annotation Search — Annotations for video or picture can even embody the spatial data. For instance a bounding field which defines the placement of the labeled object within the annotation.
  • Temporal and Spatial Search — Annotation for video can have each time vary and spatial coordinates. Therefore, we help queries which may search annotations inside the supplied time vary and spatial coordinates vary.
  • Semantics Search — Annotations may be searched after understanding the intent of the consumer supplied question. Any such search offers outcomes based mostly on the conceptually comparable matches to the textual content within the question, in contrast to the standard tag based mostly search which is anticipated to be actual key phrase matches with the annotation labels. ML algorithms additionally ingest annotations with vectors as an alternative of precise labels to help this kind of search. Person supplied textual content is transformed right into a vector utilizing the identical ML mannequin, after which search is carried out with the transformed text-to-vector to search out the closest vectors with the searched vector. Primarily based on the purchasers suggestions, such searches present extra related outcomes and don’t return empty ends in case there are not any annotations which precisely match to the consumer supplied question labels. We help semantic search utilizing Open Distro for ElasticSearch . We’ll cowl extra particulars on Semantic Search help in a future weblog article.
Semantic search
  • Vary Intersection — We lately began supporting the vary intersection queries throughout a number of annotation varieties for a particular title in the true time. This permits the purchasers to go looking with a number of knowledge labels (resulted from completely different algorithms so they’re completely different annotation varieties) inside video particular time vary or the entire video, and get the listing of time ranges or frames the place the supplied set of knowledge labels are current. A standard instance of this question is to search out the `James within the indoor shot consuming wine`. For such queries, the question processor finds the outcomes of each knowledge labels (James, Indoor shot) and vector search (consuming wine); after which finds the intersection of ensuing frames in-memory.

Search Latency

Our consumer functions are studio UI functions so that they anticipate low latency for the search queries. As highlighted above, we help such queries utilizing Elasticsearch. To maintain the latency low, now we have to ensure that all of the annotation indices are balanced, and hotspot isn’t created with any algorithm backfill knowledge ingestion for the older motion pictures. We adopted the rollover indices technique to keep away from such hotspots (as described in our blog for asset administration utility) within the cluster which may trigger spikes within the cpu utilization and decelerate the question response. Search latency for the generic textual content queries are in milliseconds. Semantic search queries have comparatively increased latency than generic textual content searches. Following graph reveals the typical search latency for generic search and semantic search (together with KNN and ANN search) latencies.

Common search latency
Semantic search latency


One of many key challenges whereas designing the annotation service is to deal with the scaling necessities with the rising Netflix film catalog and ML algorithms. Video content material evaluation performs a vital position within the utilization of the content material throughout the studio functions within the film manufacturing or promotion. We anticipate the algorithm varieties to develop extensively within the coming years. With the rising variety of annotations and its utilization throughout the studio functions, prioritizing scalability turns into important.

Knowledge ingestions from the ML knowledge pipelines are usually in bulk particularly when a brand new algorithm is designed and annotations are generated for the complete catalog. We have now arrange a distinct stack (fleet of situations) to manage the information ingestion movement and therefore present constant search latency to our customers. On this stack, we’re controlling the write throughput to our backend databases utilizing Java threadpool configurations.

Cassandra and Elasticsearch backend databases help horizontal scaling of the service with rising knowledge dimension and queries. We began with a 12 nodes cassandra cluster, and scaled as much as 24 nodes to help present knowledge dimension. This 12 months, annotations are added roughly for the Netflix full catalog. Some titles have greater than 3M annotations (most of them are associated to subtitles). At present the service has round 1.9 billion annotations with knowledge dimension of two.6TB.


Annotations may be searched in bulk throughout a number of annotation varieties to construct knowledge information for a title or throughout a number of titles. For such use circumstances, we persist all of the annotation knowledge in iceberg tables in order that annotations may be queried in bulk with completely different dimensions with out impacting the true time functions CRUD operations latency.

One of many frequent use circumstances is when the media algorithm groups learn subtitle knowledge in several languages (annotations containing subtitles on a per body foundation) in bulk in order that they will refine the ML fashions they’ve created.

Future work

There’s lots of attention-grabbing future work on this space.

  1. Our knowledge footprint retains growing with time. A number of occasions now we have knowledge from algorithms that are revised and annotations associated to the brand new model are extra correct and in-use. So we have to do cleanups for giant quantities of knowledge with out affecting the service.
  2. Intersection queries over a big scale of knowledge and returning outcomes with low latency is an space the place we wish to make investments extra time.


Burak Bacioglu and different members of the Asset Administration Platform contributed within the design and growth of Marken.