Fabricator: End-to-End Declarative Feature Engineering Platform

At Doordash, the last year has seen a surge in applications of machine learning to various product verticals in our growing business. However, with this growth, our data scientists have had increasing bottlenecks in their development cycle because of our existing feature engineering process. At a daily feature volume of over 500 unique features and 10B feature values, each component of the feature engineering process from feature generation, online materialization, offline serving, and lifecycle management was becoming operationally intensive and low velocity.

To overcome these challenges, we designed an end-to-end declarative and central feature engineering platform Fabricator. This framework leverages simple high-level YAML definitions to automate the feature pipeline orchestration using Dagster, perform scalable pipeline executions leveraging Spark on Databricks, and simplify feature store materialization and management via Redis. Additionally, the entire framework is continuously deployed, bringing iteration velocities down to just a few minutes.

In this session, we’d like to present how our Machine Learning Platform designed Fabricator by integrating various open source and enterprise solutions to deliver a declarative end-to-end feature engineering framework and take a look at the wins this enabled us to deliver. In the end, we take a closer look at key optimizations and learning and discuss plans for extending the framework for hybrid real-time and batch architectures.


Speaker

Kunal Shah

ML Platform Engineering Manager @DoorDash

Kunal Shah is an ML Platform Engineering Manager at Doordash focusing on building a feature engineering platform. Over the last year he has launched declarative frameworks for both batch and real time feature development, accelerating the development lifecycle by over 2x. Previously, he has worked on ML Platforms and Data Engineering frameworks at Airbnb and YouTube. He finished his Compute Science undergraduate at IIT Bombay, and holds a Masters in Data Science from UC Berkeley.

Read more

Date

Monday Oct 24 / 11:50AM PDT ( 50 minutes)

Track

MLOps

Share

From the same track

Session

Ray: The Next Generation Compute Runtime for ML Applications

Monday Oct 24 / 10:35AM PDT

Ray is an open source project that makes it simple to scale any compute-intensive Python workload. Industry leaders like Uber, Shopify, Spotify are building their next generation ML platforms on top of Ray.

Zhe Zhang

Head of Open Source Engineering @anyscalecompute

Session

An Open Source Infrastructure for PyTorch

Monday Oct 24 / 01:40PM PDT

In this talk we’ll go over tools and techniques to deploy PyTorch in production. The PyTorch organization maintains and supports open source tools for efficient inference like pytorch/serve, job management pytorch/torchx and streaming datasets like pytorch/data.

Mark Saroufim

Applied AI Engineer @Meta

Session

Metrics for MLOps Platforms

Monday Oct 24 / 02:55PM PDT

Many companies are investing heavily into their ML platforms, either building something in-house or working with vendors. How do we know that an ML platform is any good? How do we compare different platforms?

Chip Huyen

Co-founder @Claypot AI

Session

Empower Your ML Models with Customers Voice

Monday Oct 24 / 04:10PM PDT

ML engineers use A/B testings to iterate ML models, however, there are limitations of A/B testing that might not give us all the answers, and A/B testing might limit innovation if not used correctly.  I’ll share examples from my previous examples and lessons I learned from interviewing 10+ ML eng

Daliana Liu

Senior Data Scientist @Predibase and “The Data Scientist Show" Podcast Host

Session

Unconference: MLOps

Monday Oct 24 / 05:25PM PDT

What is an unconference? At QConLondon, we’ll have unconferences in most of our tracks.

Shane Hastie

Director of Agile Learning Programs @ICAgile