Azure Machine Learning Updates: January 2024 New Features

Azure ML’s new feature store and Prompt Flow integration changed how I structure ML pipelines in early 2024. Below are the updates that matter operationally and examples of how I used them.

Key Updates

1. Managed Feature Store GA

Feature stores enable feature reuse across ML projects:

from azure.ai.ml import MLClient
from azure.ai.ml.entities import FeatureStore, FeatureSet

# Create feature store
feature_store = FeatureStore(
    name="my-feature-store",
    description="Centralized feature repository"
)
ml_client.feature_stores.begin_create_or_update(feature_store).result()

# Define feature set
feature_set = FeatureSet(
    name="customer-features",
    version="1",
    entities=["customer_id"],
    source={
        "type": "parquet",
        "path": "azureml://datastores/features/paths/customer_features/"
    },
    features=[
        {"name": "total_purchases", "type": "float"},
        {"name": "days_since_last_purchase", "type": "int"},
        {"name": "customer_segment", "type": "string"}
    ]
)

ml_client.feature_sets.begin_create_or_update(feature_set).result()

2. Prompt Flow Integration

Build LLM applications directly in Azure ML:

# flow.dag.yaml
inputs:
  question:
    type: string
    default: "What is Azure ML?"

outputs:
  answer:
    type: string
    reference: ${llm_response.output}

nodes:
  - name: retrieve_context
    type: python
    source:
      type: code
      path: retrieve.py
    inputs:
      query: ${inputs.question}

  - name: llm_response
    type: llm
    source:
      type: code
      path: llm_call.py
    inputs:
      context: ${retrieve_context.output}
      question: ${inputs.question}
    connection: azure_openai_connection
    api: chat

3. Model Catalog Enhancements

Access and deploy foundation models:

from azure.ai.ml import MLClient

# List available models
models = ml_client.models.list(registry_name="azureml")

for model in models:
    if "llama" in model.name.lower():
        print(f"{model.name}: {model.description}")

# Deploy from catalog
deployment = ml_client.online_deployments.begin_create_or_update(
    deployment=ManagedOnlineDeployment(
        name="llama2-deployment",
        endpoint_name="llama2-endpoint",
        model="azureml://registries/azureml/models/Llama-2-7b/versions/1",
        instance_type="Standard_NC24ads_A100_v4",
        instance_count=1
    )
).result()

4. Managed Endpoints Improvements

from azure.ai.ml.entities import (
    ManagedOnlineEndpoint,
    ManagedOnlineDeployment
)

# Create endpoint with autoscaling
endpoint = ManagedOnlineEndpoint(
    name="my-endpoint",
    auth_mode="key"
)

deployment = ManagedOnlineDeployment(
    name="blue",
    endpoint_name="my-endpoint",
    model=registered_model,
    instance_type="Standard_DS3_v2",
    instance_count=1,
    scale_settings={
        "scale_type": "target_utilization",
        "min_instances": 1,
        "max_instances": 5,
        "target_utilization_percentage": 70,
        "polling_interval": "PT1M",
        "cooldown_period": "PT5M"
    }
)

MLOps Pipeline

from azure.ai.ml import dsl, Input, Output
from azure.ai.ml.entities import Pipeline

@dsl.pipeline(
    name="training-pipeline",
    description="End-to-end ML training pipeline"
)
def create_training_pipeline(
    training_data: Input,
    test_data: Input
):
    # Data preparation
    prep_job = data_prep_component(
        input_data=training_data
    )

    # Feature engineering
    feature_job = feature_engineering_component(
        input_data=prep_job.outputs.output_data
    )

    # Training
    train_job = training_component(
        training_data=feature_job.outputs.features,
        epochs=10,
        learning_rate=0.001
    )

    # Evaluation
    eval_job = evaluation_component(
        model=train_job.outputs.model,
        test_data=test_data
    )

    # Register if metrics pass
    register_job = model_registration_component(
        model=train_job.outputs.model,
        metrics=eval_job.outputs.metrics,
        min_accuracy=0.85
    )

    return {
        "model": register_job.outputs.registered_model,
        "metrics": eval_job.outputs.metrics
    }

# Create and submit pipeline
pipeline = create_training_pipeline(
    training_data=Input(type="uri_folder", path="azureml://datastores/data/paths/train/"),
    test_data=Input(type="uri_folder", path="azureml://datastores/data/paths/test/")
)

ml_client.jobs.create_or_update(pipeline)

Responsible AI Integration

from azure.ai.ml.entities import (
    ResponsibleAIInsights,
    ResponsibleAIComponentConfig
)

# Add RAI dashboard to pipeline
rai_config = ResponsibleAIComponentConfig(
    components=[
        "ErrorAnalysis",
        "Explanations",
        "Fairness",
        "Counterfactuals"
    ],
    target_column="label",
    sensitive_features=["gender", "age_group"]
)

rai_job = ResponsibleAIInsights(
    name="model-rai-analysis",
    model=trained_model,
    test_data=test_data,
    components=rai_config
)

Best Practices

1. Use feature stores for feature reuse and consistency
2. Implement MLOps pipelines for reproducibility
3. Enable autoscaling for production endpoints
4. Add RAI dashboards for model transparency
5. Version everything - data, models, and code

Conclusion

Azure ML’s 2024 updates focus on simplifying the end-to-end ML lifecycle. Feature stores, prompt flow integration, and improved model catalog make it easier to build and deploy ML solutions at scale.