Version:

Feature Views

Step-by-step tutorial on creating an feature view models.

Once you have done identity stitching to unify the identity of your users across all the cross-platforms, you can evaluate and maintain the required features/traits for each identified user using a feature views model.

Prerequisites

  • A basic Profile Builder project by following the Profile Builder CLI steps.
  • Structure of a Profile Builder project and the parameters used in different files.
  • Identity Stitching model as Feature Views reuses its output to extract the required features/traits.

Feature Views model

You can define and extract the required features/traits for an entity from your data warehouse using the feature views model. Once done, you can send them to the downstream destinations. A destination could either be the Activation API or any Reverse ETL destination that RudderStack supports. Each such destination requires data in the form of a table with an ID column and one or more feature columns.

You can use the Feature Views model to access the entity features based on any ID type and create a view having all or a specified set of entity features across the project. It also lets you unify the traits/features (defined using entity_vars) and ML models to generate a comprehensive customer 360 table.

To create a feature views model, you can add feature_views section under entities and provide a list of ID types under the id_served field. RudderStack assigns a default name to the model, if not provided, and adds all the available features on the entity into the view by default.

Default feature views model

The pb_project.yaml file for a default feature views model:

...
entities:
  - name: user
    id_types:
      - main_id
      - user_id
    feature_views:
      using_ids:
        - id: email
          name: features_by_email
        - id: salesforce_id
          name: salesforce_id_stitched_features
      features:
        - from: models/feature_table_1 #include everything from `feature_table_1` (default behaviour if `include` is not specified)
          include:	["*"]
        - from: models/feature_table_2 #exclude 'middle_name' feature from `feature_table_2`
          exclude:	["middle_name"] 

Custom feature views model

You can also define a custom feature views model by including/excluding features from any other model and adding their references to the feature_views section.

The models/profiles.yaml file for a custom feature views model:

models:
  - name: cart_feature_views
    model_type: feature_views
    model_spec:
      validity_time: 24h # 1 day
      entity_key: user
      id_served: user_id
      feature_list:
        - from: packages/pkg/models/cart_table # a table created by package
          include: ["*"] # will include all the traits
        - from: models/user_var_table
          include: ["*"]
          exclude: [cart_quantity, purchase_status] # except two, all the other traits will be included
        - from: models/sql_model
          include: [lifetime_value] # will include only one trait

Sample project

This sample project uses the output of an identity stitching model as an input to create a feature views. The following sections describe how to define your PB project files:

Project detail

The pb_project.yaml file defines the project details such as name, schema version, connection name and the entities which represent different identifiers.

You can define all the identifiers from different input sources you want to stitch together as a user_main_id:

warning
You need to add main_id to the list only if you have defined main_id_type: main_id in the ID stitcher spec.
# Project name
name: sample_id_stitching
# Project's yaml schema version
schema_version: 63
# Warehouse connection
connection: test
# Folder containing models
model_folders:
  - models
# Entities in this project and their ids.
entities:
  - name: user
    id_types:
      - main_id # You need to add `main_id` to the list only if you have defined `main_id_type: main_id` in the id stitcher spec.
      - user_id # one of the identifier from your data source.
      - email
# lib packages can be imported in project signifying that this project inherits its properties from there
packages:
  - name: corelib
    url: "https://github.com/rudderlabs/profiles-corelib/tag/schema_{{best_schema_version}}"
    # if required then you can extend the package definition such as for ID types.

Input

The input file file includes the input table references and corresponding SQL for the above-mentioned entities:

inputs:
- name: rsIdentifies
  contract: # constraints that a model adheres to
    is_optional: false
    is_event_stream: true
    with_entity_ids:
      - user
    with_columns:
      - name: timestamp
      - name: user_id
      - name: anonymous_id
      - name: email
  app_defaults:
    table: rudder_events_production.web.identifies # one of the WH table RudderStack generates when processing identify or track events.
    occurred_at_col: timestamp
    ids:
      - select: "user_id" # kind of identity sql to pick this column from above table.
        type: user_id
        entity: user # as defined in project file
        to_default_stitcher: true
      - select: "anonymous_id"
        type: anonymous_id
        entity: user
        to_default_stitcher: true
      - select: "lower(email)" # can use sql.
        type: email
        entity: user
        to_default_stitcher: true
- name: rsTracks
  contract:
    is_optional: false
    is_event_stream: true
    with_entity_ids:
      - user
    with_columns:
      - name: timestamp
      - name: user_id
      - name: anonymous_id
  app_defaults:
    table: rudder_events_production.web.tracks # another table in WH maintained by RudderStack processing track events.
    occurred_at_col: timestamp
    ids:
      - select: "user_id"
        type: user_id
        entity: user
        to_default_stitcher: true
      - select: "anonymous_id"
        type: anonymous_id
        entity: user
        to_default_stitcher: true

Model

The feature views model lets you define and extract the features/traits from your warehouse tables. Each feature is defined using an entity_var.

A sample profiles.yaml file specifying a feature views model:

var_groups:
  - name: first_group
    entity_key: user
    vars:
      - entity_var:
          name: first_seen
          select: min(timestamp::date)
          from: inputs/rsTracks
          where: properties_country is not null and properties_country != ''
      - entity_var:
          name: last_seen
          select: max(timestamp::date)
          from: inputs/rsTracks
          is_feature: false # Specifies the entity_var is not a feature
      - entity_var:
          name: user_lifespan
          select: '{{user.Var("last_seen")}} - {{user.Var("first_seen")}}'
          description: Life Time Value of a customer
      - entity_var:
          name: days_active
          select: count(distinct timestamp::date)
          from: inputs/rsTracks
          description: No. of days a customer was active
      - entity_var:
          name: campaign_source
          default: "'organic'"
      - entity_var:
          name: user_rank
          default: -1
      - entity_var:
          name: campaign_source_first_touch
          select: first_value(context_campaign_source)
          window:
              order_by:
                  - timestamp asc
              partition_by:
                  - main_id
          from: inputs/rsIdentifies
          where: context_campaign_source is not null and context_campaign_source != ''
      - input_var:
          name: num_c_rank_num_b_partition
          select: rank()
          from: inputs/tbl_c
          default: -1
          window:
            partition_by:
              - '{{tbl_c}}.num_b'
            order_by:
              - '{{tbl_c}}.num_c asc'
          where: '{{tbl_c}}.num_b >= 10'
      - entity_var:
          name: min_num_c_rank_num_b_partition
          select: min(num_c_rank_num_b_partition)
          from: inputs/tbl_c

var_groups

The var_groups field groups all the vars under it and provides the provision to define any configuration keys that need to be shared across vars.

FieldData typeDescription
nameStringName to identify the var_groups uniquely.
entity_keyStringSpecifies the entity to be used.
varsListSpecifies the entity_var and input_var variables.

entity_var

The entity_var field provides inputs for the feature views model. This variable stores the data temporarily, however, you can choose to store its data permanently by specifying the name in it as a feature in the features key.

FieldData typeDescription
nameStringName of the entity_var to identify it uniquely.
selectStringColumn name/value you want to select from the table. This defines the actual value that will be stored in the variable. You can use simple SQL expressions or select an entity_var as {{entityName.Var(\"entity_var\")}}. It has to be an aggregate operation that ensures the output is a unique value for a given main_id. For example: min(timestamp), count(*), sum(amount) etc. This holds true even when a window function (optional) is used. For example:: first_value(), last_value() etc are valid while rank(), row_number(), etc. are not valid and give unpredictable results.
fromListReference to the source table from where data is to be fetched. You can either refer to another model from the same YAML or some other table specified in input YAML.
whereStringAny filters you want to apply on the input table before selecting a value. This must be SQL compatible and should consider the data type of the table.
defaultStringDefault value in case no data matches the filter. When defining default values, make sure you enclose the string values in single quotes followed by double quotes to avoid SQL failure. However, you can use the non-string values without any quotes.
descriptionStringTextual description of the entity_var.
is_featureBooleanDetermines whether the entity_var is a feature. The default value is true.
windowObjectSpecifies the window function. Window functions in SQL usually have both partition_by and order_by properties. But for entity_var, partition_by is added with main_id as default; so, adding partition_by manually is not supported. If you need partitioning on other columns too, check out input_var where partition_by on arbitrary and multiple columns is supported.

input_var

The syntax of input_var is similar to entity_var, with the only difference that instead of each value being associated to a row of the feature views, it’s associated with a row of the specified input. While you can think of an entity_var as adding a helper column to the feature views, you can consider an input_var as adding a helper column to the input.

info
If more than one input_var are required to derive an entity_var, then all the input_var must be defined on the same table.
FieldData typeDescription
nameStringName to store the retrieved data.
selectStringData to be stored in the name.
fromListReference to the source table from where data is to be fetched.
whereString(Optional) Applies conditions for fetching data.
defaultString(Optional) Default value for any entity for which the calculated value would otherwise be NULL.
descriptionString(Optional) Textual description.
windowObject(Optional) Specifies a window over which the value should be calculated.

window

FieldData typeDescription
partition_byString(Optional) List of SQL expressions to use in partitioning the data.
order_byString(Optional) List of SQL expressions to use in ordering the data.

In window option, main_id is not added by default, it can be any arbitrary list of columns from the input table. So if a feature should be partitioned by main_id, you must add it in the partition_by key.

Output

After running the project, you can view the generated material tables.

A sample output containing the results in Snowflake:

Generated table (Snowflake)

Questions? Contact us by email or on Slack