Azure Machine Learning Feature Store Python SDK

The azureml-featurestore package is the core SDK interface for Azure ML Feature Store. This SDK works along the azure-ai-ml SDK to provide the managed feature store experience.

Main features in the azureml-featurestore package

• Develop feature set specification in Spark with the ability for feature transformation.
• List and get feature sets defined in Azure ML Feature Store.
• Generate and resolve feature retrieval specification.
• Run offline feature retrieval with point-in-time join.

Getting started

You can install the package via pip install azureml-featurestore

To learn more about Azure ML managed feature store visit https://aka.ms/featurestore-get-started

Change Log

1.2.1 (2025.09.17)

• Fix bugs

1.2.0 (2025.08.18)

• Improve online featurestore

1.1.2 (2025.03.27)

• Fix bugs

1.1.1 (2025.02.25)

• Fix logging issue

1.1.0 (2024.03.12)

New Features:

• [Public Preview] Support for DSL (Domain Specific Language) for feature definition. The DSL is a simplified way to define feature set transformations using a declarative syntax.
  • DSL feature set supports custom source
  • DSL feature set supports temporal join lookback, source delay and source lookback
  • DSL feature set supports load from materialized store
  • get_offline_features supports feature sets with different transformations(dsl, udf or none).

1.0.1 (2023.12.28)

• Update dependencies

1.0.0 (2023.11.14)

• [GA] Custom feature source: Custom feature source supports customized source process code script with a user defined dictionary as input.
• [GA] International regions and sovereign cloud support.
• [GA] Offline backfill materialization now replaces all data within a feature window instead of doing upsert based on timestamp.
• [GA] Added bootstrap option for materialization, which enables materializing data from offline store into online store.
• Re-enabling materialization in a feature set now invalidates all previously materialized data.
• Feature set spec dump now accepts a file path or a folder path as dump target, and an overwrite option to control whether to override the target.
• Various bug fixes

0.1.0b6 (2023.11.1)

• Various bug fixes

0.1.0b5 (2023.10.4)

• Various bug fixes

0.1.0b4 (2023.08.28)

New Features:

• [Public preview] Added custom feature source. Custom feature source supports customized source process code script with a user defined dictionary as input.

• [Public preview] Added csv feature source, deltatable feature source, mltable feature source, parquet feature source as new feature source experience. Previous feature source usage compatibility will be deprecated in 6 months.

• Bug fixes

Breaking changes:

• Moved init_online_lookup, shutdown_online_lookup and get_online_features out of FeatureStoreClient, and into the module as standalone functions.
• get_online_features contract changed from accepting (for the observation_data argument) and returning pandas.DataFrame to accepting (as the observation_data argument) and returning pyarrow.Table.

Other changes:

• Moved online feature store support into public preview.

0.1.0b3 (2023.07.10)

• Various bug fixes

0.1.0b2 (2023.06.13)

New Features:

• [Private preview] Added online store support. Online store supports materialization and online feature values retrieval from Redis cache for batch scoring.

• Various bug fixes

0.1.0b1 (2023.05.15)

New Features:

Initial release.