7 - Managed feature store
Adds an Azure ML managed feature store workspace. The Parquet data source and a feature set are planned as a separate feature store workload pack.
| Id | feature-store |
| Network | public |
| Serving | online |
| Adds | Managed feature store workspace |
| IaC and CI/CD | Terraform or Bicep; GitHub or Azure DevOps |
Adds a managed feature store workspace alongside the project workspace. The feature store always gets a managed network in AllowInternetOutbound mode, because serverless Spark needs one and Microsoft documents it as the only supported mode for feature stores (Tutorial 6). The workspace sets the Spark runtime to 3.4.0 and has no container registry, because a managed network requires a Premium registry and the feature store does not need one.
This scenario creates the feature store workspace (kind: FeatureStore) and nothing else. The generated training pipeline does not define, register or materialize a feature set, and does not read features from the store. The infrastructure is applied, verified and destroyed live on Terraform and Bicep by the e2e.yml workflow. The feature set workload pack is not verified live. The Microsoft tutorial Develop and register a feature set covers the missing parts: a feature set spec, a Spark transformation, an offline store, a materialization identity and role assignments.
Use it
See the cost estimates and the prerequisites.