If you’d like to follow along, here are the GitHub repositories for the Feast demo and Tecton demo.
Creating features with Feast
Feast primarily helps in serving features for online inference and offline training, but not in transforming feature values (Feast takes transformed values as input). Creating a feature in Feast consists of registering a pre-processed feature, along with its metadata.
Creating features with Tecton
When you create a feature in Tecton, you're actually defining the feature transformation itself. Tecton includes data pipeline management, stores the materialized data in online and offline stores, and runs the backfill for you.
Managing data pipelines with Feast
Feast takes transformed values as input so you'll have to manage your own data pipelines. You'll have to:
Create a streaming pipeline (e.g., with Spark)
Create a batch pipeline (e.g., with DBT)
Set up monitoring to make sure your features are properly sent to Feast and watch out for train / serve skew
Set up a backfill job with Airflow
Managing data pipelines with Tecton
Tecton handles the feature transformations for you. Tecton will create and manage data pipelines for you in the background. Tecton will also handle monitoring and alerting as soon as you create the feature. Tecton’s goal is to make it easy for teams to quickly develop real-time data pipelines for ML applications.
Serving features with Feast
Both Feast and Tecton allow you to serve features online and create training datasets. They share the same feature retrieval APIs:
'get_offline_features' to generate a training dataset with point-in-time correct features with labels
'get_online_features' to retrieve the features you need to make a prediction
Serving features with Tecton
Tecton allows you to serve features online and create training datasets. Tecton has the same feature retrieval APIs as Feast. The main difference is that Tecton manages the online store for you, which includes a layer of monitoring and alerting (for example, if feature values stop coming through and your feature is getting stale, you'll receive an alert that your upstream data source might be down).
Scaling Considerations
Feast requires manual, custom configuration to power production ML at scale. Common considerations include how to:
Reliably and scalably compute batch and / or streaming features
Reliably and scalably backfill features to the offline / online store
Manage online store costs given large data volumes
Tecton is a fully hosted solution that requires no manual configuration and ensures enterprise scalable machine learning and reliability, with SLA guarantees.
Recap
Use Feast if you:
Need to deploy on-prem, on GCP, or on Azure
Have the resources and skills to manage your own feature store
Need a highly customizable solution
Don’t need any help with streaming or real-time data pipelines
Use Tecton if you:
Want minimal engineering overhead in managing your own feature store
Need mission-critical reliability, scalability, and/or support
Want to automate batch, streaming, and real-time features
