Azure Machine Learning Feature Store SDK

1.2.2 · active · verified Sun Apr 12

The `azureml-featurestore` package is the core SDK interface for Azure ML Feature Store, working alongside `azure-ai-ml` to provide a managed feature store experience. It enables the development of feature set specifications in Spark, listing and retrieving feature sets, generating and resolving feature retrieval specifications, and performing offline feature retrieval with point-in-time joins. The library is actively developed, with its current version being 1.2.2.

Warnings

breaking The online feature retrieval functions `init_online_lookup`, `shutdown_online_lookup`, and `get_online_features` were moved from being methods of `FeatureStoreClient` to standalone module-level functions. Additionally, the contract for `get_online_features` changed from accepting/returning `pandas.DataFrame` to `pyarrow.Table`.
Fix: Update your code to import these functions directly from `azureml.featurestore` (e.g., `from azureml.featurestore import get_online_features`) and ensure your data handling for `get_online_features` uses `pyarrow.Table`.
gotcha Many operations with Azure ML Feature Store, including creating the feature store itself, retrieving features, and materialization jobs, require specific Azure Role-Based Access Control (RBAC) permissions. Common errors arise from insufficient permissions.
Fix: Ensure the user identity or managed identity running the operations has the `AzureML Data Scientist` role on the feature store and `Storage Blob Data Reader` on relevant source and offline store storage accounts. Consult Azure ML documentation for detailed RBAC requirements for each operation.
gotcha Updating an offline materialization store at the feature store level requires disabling offline materialization for all feature sets within that store first. Disabling materialization resets the status of already materialized data, rendering it unusable.
Fix: Before updating the offline materialization store, disable offline materialization on all affected feature sets. After the store update, you must re-enable materialization and resubmit all materialization (backfill) jobs to regenerate the data.

Install

pip install azureml-featurestore azure-ai-ml mltable Install core SDK and dependencies

Imports

FeatureStoreClient

from azureml.featurestore import FeatureStoreClient

init_online_lookup
```
from azureml.featurestore import init_online_lookup
```
As of breaking changes around version 0.1.0b3, `init_online_lookup` was moved from a method of `FeatureStoreClient` to a standalone function in the module.
shutdown_online_lookup
```
from azureml.featurestore import shutdown_online_lookup
```
As of breaking changes around version 0.1.0b3, `shutdown_online_lookup` was moved from a method of `FeatureStoreClient` to a standalone function in the module.
get_online_features
```
from azureml.featurestore import get_online_features
```
As of breaking changes around version 0.1.0b3, `get_online_features` was moved from a method of `FeatureStoreClient` to a standalone function in the module.

Quickstart

This quickstart demonstrates how to initialize the `MLClient` for managing Azure ML resources (including creating a Feature Store if needed) and then the `FeatureStoreClient` from `azureml-featurestore` for interacting with feature sets. It includes a placeholder for listing feature sets to show basic client usage. Authentication uses `DefaultAzureCredential` and expects Azure environment variables or hardcoded values.

import os
from azure.ai.ml import MLClient
from azure.ai.ml.entities import FeatureStore, FeatureStoreEntity
from azure.identity import DefaultAzureCredential
from azureml.featurestore import FeatureStoreClient

# Replace with your actual subscription, resource group, and feature store name
subscription_id = os.environ.get('AZURE_SUBSCRIPTION_ID', 'your-subscription-id')
resource_group_name = os.environ.get('AZURE_RESOURCE_GROUP', 'your-resource-group')
feature_store_name = os.environ.get('AZURE_FEATURE_STORE_NAME', 'my-feature-store')
feature_store_location = os.environ.get('AZURE_LOCATION', 'eastus')

# Authenticate and create MLClient for managing Azure ML resources
try:
    credential = DefaultAzureCredential()
    ml_client = MLClient(
        credential=credential,
        subscription_id=subscription_id,
        resource_group_name=resource_group_name
    )
except Exception as e:
    print(f"Could not authenticate or create MLClient: {e}")
    print("Please ensure AZURE_SUBSCRIPTION_ID, AZURE_RESOURCE_GROUP, and AZURE_LOCATION environment variables are set or replaced.")
    exit(1)

# Create a Feature Store if it doesn't exist (this step typically requires Azure CLI or a separate script to run once)
# For a runnable quickstart, we assume the feature store might already be created or we just define it.
# In a real scenario, you'd check for existence and create if necessary.
print(f"Attempting to get or define Feature Store '{feature_store_name}'...")
feature_store = FeatureStore(
    name=feature_store_name,
    location=feature_store_location,
    description="My first Azure ML Feature Store"
)

# This line would typically be ml_client.feature_stores.begin_create_or_update(feature_store).result()
# However, for a quickstart that focuses on the azureml-featurestore SDK, we will mock the client initialization
# assuming the feature store exists.
print("Assuming feature store is created. Initializing FeatureStoreClient...")

# Initialize FeatureStoreClient
# This client is used for developing and consuming features.
fs_client = FeatureStoreClient(
    credential=credential,
    subscription_id=subscription_id,
    resource_group_name=resource_group_name,
    name=feature_store_name
)

print(f"Successfully initialized FeatureStoreClient for '{feature_store_name}'.")

# Example: List existing feature sets (requires a feature store to be set up and contain feature sets)
# In a real scenario, you would have feature sets defined and registered.
# This is a placeholder to demonstrate client usage.
print("Attempting to list feature sets (this might return empty if no feature sets exist)...")
feature_sets = fs_client.feature_sets.list()
for fs in feature_sets:
    print(f"  - Found Feature Set: {fs.name} (Version: {fs.version})")

print("Quickstart finished. For advanced usage, refer to Azure ML Feature Store documentation.")

view raw JSON →