blurr-dev 0.757

Data aggregation pipeline for running real-time predictive models

What is Blurr?

Blurr transforms structured, streaming raw data into features for model training and prediction using a high-level expressive YAML-based language called the Data Transform Configuration (DTC).

The motivation for building Blurr is to help data scientists extract features from raw data without depending on a data engineer. Reducing engineering dependencies makes iterations on feature engineering and selection faster than existing methods - writing Spark code, using SparkSQL or KSQL. We wanted to build something that is portable across any infrastructure stack.

Blurr has the same workflow during model development and taking it to production, so productionizing an ML pipeline is a breeze.

Blurr vs. stream/batch processors

The DTC is a data transform definition for structured data. The DTC encapsulates the business logic of data transforms and Blurr orchestrates the execution of data transforms.

Blurr is processor-agnostic, so DTCs can be run by event processors such as Spark.

1. Blurr is to Spark as Hibernate is to databases
2. Blurr can be used on Spark in the same way as SparkSQL
3. Blurr is WORAIS (Write Once, Run on Any Infrastructure Stack)

Because real world infrastructure is extremely diverse, Blurr is designed to run on virtually any infrastructure stack that runs Python 3.6+.

Give us feedback on the metaphors and help improve Blurr!

The future of MLOps

We believe in a world where everyone is a data engineer. Or a data scientist. Or an ML engineer. The lines are blurred (cough). Just like development and operations became DevOps over time

--- Blurr authors

Blurr is a collection of components built for MLOps, the DTC is one of them. DTC ⊆ Blurr

We see a future where MLOps means teams putting together various technologies to suit their needs. For production ML applications, the speed of experimentation and iterations is the difference between success and failure. The DTC helps teams iterate on features faster. The vision for Blurr is to build MLOps components to help ML teams experiment at high speed.

Table of contents

• DTC at a glance
• Tutorial & Docs
• Install
• Contribute
• Data Science 'Joel Test'
• Roadmap

Coming up with features is difficult, time-consuming, requires expert knowledge. 'Applied machine learning' is basically feature engineering

--- Andrew Ng

DTC at a glance

Raw data like this

{ "user_id": "09C1", "session_id": "915D", "country" : "US", "event_id": "game_start" }
{ "user_id": "09C1", "session_id": "915D", "country" : "US", "event_id": "game_end", "won": 1 }
{ "user_id": "09C1", "session_id": "915D", "country" : "US", "event_id": "game_start" }
{ "user_id": "09C1", "session_id": "915D", "country" : "US", "event_id": "game_end", "won": 1 }
{ "user_id": "B6FA", "session_id": "D043", "country" : "US", "event_id": "game_start" }
{ "user_id": "B6FA", "session_id": "D043", "country" : "US", "event_id": "game_end", "won": 1 }
{ "user_id": "09C1", "session_id": "T8KA", "country" : "UK", "event_id": "game_start" }
{ "user_id": "09C1", "session_id": "T8KA", "country" : "UK", "event_id": "game_end", "won": 1 }

turns into

session_id	user_id	games_played	games_won
915D	09C1	2	2
D043	B6FA	1	1
T8KA	09C1	1	1

using this DTC

Type: Blurr:Transform:Streaming
Version: '2018-03-01'
Name : sessions

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

Identity: source.user_id

Time: parser.parse(source.timestamp)

Aggregates:

 - Type: Blurr:Aggregate:BlockAggregate
   Name: session_stats
   Store: hello_world_store

   Split: source.session_id != session_stats.session_id

   Fields:

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

     - Name: games_played
       Type: integer
       When: source.event_id == 'game_start'
       Value: session_stats.games_played + 1

     - Name: games_won
       Type: integer
       When: source.event_id == 'game_end' and source.won == '1'
       Value: session_stats.games_won + 1

Tutorial and Docs

Read the docs

Streaming DTC Tutorial | Window DTC Tutorial

Preparing data for specific use cases using Blurr

Dynamic in-game offers (Offer AI) | Frequently Bought Together

Use Blurr

We interact with Blurr using a Command Line Interface (CLI). Blurr is installed via pip:

$ pip install blurr

Transform data

$ blurr transform \
     --streaming-dtc ./dtcs/sessionize-dtc.yml \
     --window-dtc ./dtcs/windowing-dtc.yml \
     --source file://path

CLI documentation

Contribute to Blurr

Welcome to the Blurr community! We are so glad that you share our passion for making data management and machine learning accessible to everyone.

Please create a new issue to begin a discussion. Alternatively, feel free to pick up an existing issue!

Please sign the Contributor License Agreement before raising a pull request.

Data Science 'Joel Test'

Inspired by the (old school) Joel Test to rate software teams, here's our version for data science teams. What's your score? We'd love to know!

1. Data pipelines are versioned and reproducible
2. Pipelines (re)build in one step
3. Deploying to production needs minimal engineering help
4. Successful ML is a long game. You play it like it is
5. Kaizen. Experimentation and iterations are a way of life

Roadmap

Blurr is currently in developer preview. Stay in touch!: Star this project or email [e-mail]

• Local transformations only
• Support for custom functions and other python libraries in the DTC
• Spark runner
• S3-S3 data transformations
• Add DynamoDB as a Store
• Features server