The core of Blurr is the Blurr Transform Spec (BTS).

There are 2 types of BTSs - Streaming and Window BTS.

The Streaming BTS takes raw data and converts it into blocks.

The Window BTS generates data relative to a block. This is especially useful for building features for machine learning models.

Blurr is in Developer Preview so the spec is likely to change

The anatomy of a Streaming BTS

The raw event being processed is defined as source in the BTS. Event parameters are accessed as source.<key> such as source.user_id. BTSs are processed in a python 3.6 environment and the expressions used for field values are executed as python statements.

Header

Type: Blurr:Transform:Streaming
Version: '2018-03-07'
Description: Create blocks from streaming Raw Data
Name: offer_ai_v1
Identity: source.user_id
Time: parse_time(source.event_time, 'YYYY/mm/dd HH:MM:SS')
When: source.package_version = '1.0'

Key	Description	Allowed values	Required
Type	The type of BTS - Streaming or Window	Blurr:Transform:Streaming or Blurr:Transform:Window	Required
Version	Version number of the BTS, used by the Blurr Transform Library to parse the template	Specific BTS versions only 2018-03-01	Required
Description	Text description for the BTS	Any string	Optional
Name	Unique name for the BTS	Any string	Required
Identity	The dimension in the raw data around which data is aggregated	source.<field>	Required
Time	Define what field in the source data to use as the time of the event. This is available as time for other expressions to use. If source does not contain time then the datetime.utcnow() expression can be used to retrieve the current utc datetime	Any python datetime object	Optional
When	Boolean expression that defines which raw data should be processed	Any boolean expression	Optional

Store

At the end of a transform, data is persisted in the store. The Blurr Transform Library works with an abstraction of storage which isolates the actual storage tier from how it works with the DTL.

Stores:
   - Type: Blurr:Store:Memory
     Name: hello_world_store

Key	Description	Allowed values	Required
Type	The destination data store	Blurr:Store:Memory. More Stores such as S3 and DynamoDB coming soon	Required
Name	Name of the store, used for internal referencing within the BTS	Any string	Required

Import

The import section can be used to specify arbitrary python code that can be used in the BTS.

As shown below Time key is using datetime and timezone and these are being specified in the import section.

Import:
  - { Module: datetime, Identifiers: [ datetime, timezone ]}

Time: datetime.fromtimestamp(source.timestamp, timezone.utc)

Similarly any custom python code written can be incorporated into the BTS:

Import:
  - { Module: my_module, Identifiers: [ my_function1, my_function2 ]}

After this my_function1 and my_function2 can be used in the BTS. This is equivalent to a from my_module import my_function1, my_fuction2 python statement.

Key	Description	Allowed values	Required
Module	Python Module to import	Any module that is available to the python execution runtime	Required
Identifiers	List of identifiers to import	The identifiers provided should be available in the module. If not identifiers are provided then the module is imported directly.	Optional

Some examples and its equivalent python statement

Import	Python statement
Module: my_module, Identifiers: [ my_function1, my_function2 ]	from my_module import my_function1, my_fuction2
Module: my_module	import my_module
Module: my_module as mod_name	import my_module as mod_name
Module: my_module, Identifiers: [ my_function1 as func1 ]	from my_module import my_function1 as func1

Aggregates

Aggregates defines groups of data that are either in a one-to-one relationship with the Identity or a in a many-to-one relationship.

There are 3 types of Aggregates in a Streaming BTS.

Identity Aggregate

Blurr:Aggregate:Identity. Fields in the Identity Aggregates are in a one-to-one relationship with the identity. There is a single record that stores these fields and change to these fields overwrite the previous value. There are no historical records kept for state changes.

Aggregates:

  - Type: Blurr:Aggregate:Identity
    Name: user
    Store: offer_ai_dynamo
    When: source.event_id in ['app_launched', 'user_updated']

    Fields:

      - Name: user_id
        Type: string
        Value: source.customer_identifier

      - Name: country
        Value: source.country

Key	Description	Allowed values	Required
Type	Type of Aggregate	Blurr:Aggregate:Identity, Blurr:Aggregate:Block, Blurr:Aggregate:Variable	Required
Name	Name of the Aggregate	Any string that does not start with _ or run_. , unique within the BTS	Required
Store	Name of the Store in which to create the Aggregate	Stores defined in the BTS	Required
When	Boolean expression that defines which raw events to process	Any boolean expression	Optional

An Aggregate contains fields for the information being stored. Identity Aggregate fields are especially useful for data that is relatively static over time - like a user's country.

Each field in an Aggregate has 4 properties.

Key	Description	Allowed values	Required
Name	Name of the field	Any string that does not start with _ or run_.	Required
Type	Type of data being stored	integer, boolean, string, datetime, float, map, list, set	Required
Value	Value of the field	Any python expression, and must match the Type	Required
When	Boolean expression that defines which raw events to process	Any boolean expression	Optional

All fields in the Aggregate are encapsulated in a Aggregate object. The object is available in the DTL, which is the python environment processing the BTS. Field values can be accessed using AggregateName.FieldName

An Identity Aggregate can also have Dimensions to split the aggregation along those dimensions. Dimensions are also a list of Fields with only integer, boolean and string types supported. Each raw event that is evaluated should contain the data needed to determine all the Dimensions specified.

Example:

Aggregates:

  - Type: Blurr:Aggregate:Identity
    Name: user
    Store: offer_ai_dynamo
    When: source.event_id in ['app_launched', 'user_updated']

    Dimensions:
      - Name: level
        Type: string
        Value: source.level

    Fields:

      - Name: events
        Type: int
        Value: user.events + 1

      - Name: country
        Value: source.country

Block Aggregate

Blurr:Aggregate:Block. Fields in the Block Aggregates are in a one-to-many relationship with the identity. These fields are aggregated together in blocks based on the Dimensions fields specified. Blurr:Aggregate:Block is similar to Blurr:Aggregate:Identity with the difference being that in the case of Block aggregates Blurr internally keeps track on the time-range of the block and Blurr:Aggregate:Block aggregates are expected to be spread across time.

- Type: Blurr:Aggregate:Block
  Name: session
  When: source.app_version > '3.7'

  Dimensions:
    - Name: session_id
      Type: string
      Value: source.session_id

  Fields:
    - Name: presents_wrapped
      Type: integer
      Value: session.presents_wrapped + 1
      Filter: source.event_name == 'wrap_present'

Key	Description	Allowed values	Required
Type	Type of Aggregate	Blurr:Aggregate:Identity, Blurr:Aggregate:Block, Blurr:Aggregate:Variable	Required
Name	Name of the Aggregate	Any string that does not start with _ or run_. , unique within the BTS	Required
When	Boolean expression that defines which raw events to process	Any boolean expression	Optional

Activity Aggregate

Blurr:Aggregate:Activity.Activity Aggregate is a specialization of the Block Aggregate where the inactivity based aggregate splitting condition is specified by SeparateByInactiveSeconds.

Variable

Blurr:Aggregate:Variable. Variable Aggregates are temporary variables that can be used in other data blocks. Variables aim to reduce code duplication and improve readability. They are useful for cleansing / modifying / typecasting source elements and representing complex filter conditions that evaluate to a binary value.

- Type: Blurr:Aggregate:Variable
  Name: vars
  Fields:
    - Name: item_price_micro
      Type: integer
      Value: int(float(source.item_price) * 1000000)

This can be accessed as vars.item_price_micro anywhere in the Streaming BTS

Window BTS

Windowing concepts in Blurr are similar to the concepts in Apache Spark/Flink.

Once the Streaming BTS has run, the raw data is split into blocks. The Window BTS generates data relative to a block.

Say we're training a model to predict what sales offer to show to a user in a session. We want the features of this model to be purchases_prev_week, games_played_last_session and win_ratio_last_session. We'd like to predict next_7_days_revenue.

When training a model, the data needs to be organized around the decision point. We want to know the value of the features at the point an offer is shown. So if the offer is shown during session 4, the features need to be relative to session 4. And the optimization function (total revenue over next 7 days) needs to be relative to session 4 as well.

The decision point is the Anchor. A window defines segments of data relative to the anchor. In this case, we need the features and optimization function relative to when an offer is shown.

The anatomy of a Window BTS

Header

Type: Blurr:Transform:Window
Version: '2018-03-01'
Description: Generate features around an Anchor
Name: Window Example
SourceBTS: offer_ai_v1

Key	Description	Allowed values	Required
Type	The type of BTS - Streaming or Window	Blurr:Transform:Streaming or Blurr:Transform:Window	Required
Version	Version number of the BTS, used by the Blurr Transform Library to parse the template	Specific BTS versions only 2018-03-01	Required
Description	Text description for the BTS	Any string	Optional
Name	Unique name for the Window BTS	Any string	Required
SourceBTS	The name of the streaming BTS upon which this window BTS will be executed	Valid streaming BTS	Required

Anchor

An Anchor represents a decision made at a point in time. For example: when a model to pick the offer price for a user is being used, then this decision being made is an Anchor.

This is a vital concept to align the way data processing is done for training data generation and for prediction. The training data is generated as if the only information that was available was up to the Anchor point. Anything after the Anchor point can only be used as a label (which is what the model will predict).

Anchor:
  Condition:  offer_ai_v1.game_stats.offer_type != ''
  Max: 1

Condition is a python expression which returns a boolean value to determine if an anchor condition exists in the block being processed. Field values in a block are accessed as StreamingBTSName.AggregateName.FieldName.

Max defines the maximum number of output rows to be generated every time the Window BTS is run.

When the number of blocks that satisfy the anchor condition is more than the max value specified, then, a ‘max’ number of satisfying sessions are picked randomly. Max can be used to prevent oversampling of more active users.

Store

Stores:
   - Type: Blurr:Store:Memory
     Name: hello_world_store

Key	Description	Allowed values	Required
Type	The destination data store	Blurr:Store:Memory. More Stores such as S3 and DynamoDB coming soon	Required
Name	Name of the store, used for internal referencing within the BTS	Any string	Required

Aggregates

All Aggregate operations that are performed in a window BTS can only use the following available data:

1. Anchor block - Block that satisfies the anchor condition. The fields from the anchor block can be accessed as anchor.FieldName.
2. Identity Aggregate - Identity aggregates available from the source Streaming BTS. The fields from an Identity Aggregate can be accessed as StreamingBTSName.IdentityAggregateName.Field Name.
3. A window of blocks around the anchor block - A list of blocks from a Block Aggregate before or after the anchor block based on the window defined. A field from the list of blocks is referenced as WindowName.FieldName.

Window Aggregate

Aggregates:

  Aggregates:
   - Type: Blurr:Aggregate:Window
     Name: last_session
        # Defines a processing window for the rollup. Supported window types are Day, Hour and Count
     WindowType: count
        # Negative values are backward from the Anchor
     WindowValue: -1
     Source: offer_ai_v1.game_stats

     Fields:

         # The output will contain a column for last_session.games_played
      - Name: games_played
        Type: integer
         # source.games_played is a list containing the games_played information from the
         # sessions that fall in the window defined above. In this case, because of a count
         # window of 1 the source.games_played list only contains a single data point.
        Value: source.games_played[0]

Key	Description	Allowed values	Required
Type	Type of Aggregate	Blurr:Aggregate:Window, Blurr:Aggregate:Variable	Required
Name	Name of the Aggregate	Any string, unique within the BTS	Required
WindowType	The type of window to use around the anchor block	day, hour, count	Optional. A Window Aggregate can be defined without a Window
WindowValue	The number of days, hours or blocks to window around the anchor block	Integer	Optional. A Window Aggregate can be defined without a Window
Source	The Block Aggregate or Activity Aggregate (defined in the Streaming BTS) on which the window operations should be performed	Valid Block of Activity Aggregate	Required

All functions defined on windows work on a list of values. For e.g. if a session contains a games_played field and a last_week window is defined on it, then last_week.games_played represents the list of values from last week's sessions.

Important: Window operations using Window Aggregate do not include the Anchor block itself.

Each field in a Window Aggregate has 3 properties.

Key	Description	Allowed values	Required
Name	Name of the field	Any string	Required
Type	Type of data being stored	integer, boolean, string, datetime, float, map, list, set	Required
Value	Value of the field	Any python expression, and must match the Type	Required

Variable

The Variable Aggregate works exactly the same as in the Streaming BTS.

Order of Aggregates

Aggregates can be defined in any order. However, if field values are referenced within the BTS, they must be defined in the order in which they are referenced. For example, if a Block Aggregate uses a Variable, then the Variable Aggregate should be defined before the Block Aggregate so that the Variable is processed first and available to the Block Aggregate when the Block Aggregate is being processed.

Processing field values

All field values must be valid python expressions. These expressions are executed by the Blurr Transform Library (DTL), which uses a Python 3.6 interpreter.

Errors

Field values are not recorded when:

1. Evaluation results in an error
2. A None value is returned / no values are returned
3. Value returned is of a type different from the type of the field and casting the value to the required type fails.

Types

The following types are supported.

Type	Description	Expected format	Default
integer	integers!	2	0
boolean	boolean!	true	false
string	default type	'this is a string'	''
datetime	datetime	any datetime object in python	-
float	floating point numbers with decimals	1.2	0.0
map	Any type to type maps	‘{“a”: 1, “b”: 2}’	empty map
list	lists can contain any type of data	‘[1,2,3]’	empty list
set	contains only unique elements that are integer, long, float and strings	-	empty set