We are excited to announce the public preview of Azure ML managed feature store. Managed feature store empowers machine learning professionals to develop and productionize features independently. You simply provide a feature set specification and let the system handle serving, securing, and monitoring of your features, freeing you from the overhead of setting up and managing the underlying feature engineering pipelines.
By integrating with our feature store across the machine learning life cycle, you can accelerate model experimentation, increase the reliability of your models, and reduce your operational costs. This is achieved by redefining the machine learning DevOps experience.
This is a quote from our customer:
“Azure Machine Learning managed feature store allows us to build more accurate and robust AI models, achieving unprecedented performance for money laundering and fraud detection use cases. The capability allows us to easily create, store, and access features for machine learning. We are excited to be working with Microsoft to continue to push the boundaries of AI development and look forward to the possibilities from this new innovation.”
-Nicolas Goosse | Head of Artificial Intelligence | Belfius
How feature store optimizes your team's workflow
- Search and reuse your team’s features to avoid duplicate work and deliver consistent predictions.
- Create new features with transformation abilities to address feature engineering requirements with agility.
- Empower your team to engage in deep work while feature store automatically operationalizes and manages feature engineering pipelines required for transformation and materialization.
- Use the same feature pipelines used for training data generation in inference to avoid training/serving skew.
Discover and reuse features across your organization
You can search and reuse features across feature stores. You can also use the feature store from spark environments including Azure ML workspaces and Azure Databricks.
Create features with transformations
- Support for custom transformations - Develop features with custom transformations, such as window-based aggregates by writing a Spark transformer (like the above code).
- Support for precomputed features - Bring precomputed features into feature store and serve them without writing code.
- Local development and testing - Fully develop and test feature sets locally with a Spark environment
Materialize features
Materialization is the process of computing feature values for a given feature window and persisting in a materialization store. Now feature data can be retrieved faster and more reliably for training and inference purposes.
- Managed feature materialization pipeline - Declaratively specify the materialization schedule, and the system takes care of scheduling, precomputing and materializing the values into the materialization store.
- Backfill support - Perform on-demand materialization of feature sets for a given feature window
MLOps support
- The built in feature retrieval component allows you to operationalize training and inference pipelines without writing any code for feature augmentation
- All feature store entities definitions including feature sets can be persisted as yaml files in your source repo and rolled out to the feature stores using GitOps
- Feature sets are versioned and immutable, allowing newer versions of models to use newer version of features without disrupting the SLA of the older version of the model
View lineage
For a given feature set, you can see the list of models that depend on it. For a given model, you can also see the list of feature sets it depends on.
Secure features
- RBAC - Role based access control for feature store, feature set and entities.
- Query across feature stores - You can create multiple feature stores with different access for users, but perform querying (for example, generate training data) from across multiple feature stores
Summary
Managed feature store lets your machine learning team develop and productionize features independently while making your machine learning lifecycle more agile. Take it for a spin here
Learn more
To learn more, watch the Microsoft Build 2023 sessions to get familiar with other Azure Machine Learning announcements.
- Breakout: Build and maintain your company Copilot with Azure Machine Learning and GPT-4
- Breakout: Practical deep-dive into machine learning techniques and MLOps
- Breakout: Building and using AI models responsibly
- Evaluate, finetune and deploy open source models curated by the AzureML team.
- Deploying Hugging Face Hub models in Azure Machine Learning
Microsoft