Question

Your data science team repeatedly loads a Parquet dataset stored in an Azure Blob datastore for training. Each training run needs consistent schema handling and column type parsing, and the team wants to reference the same logical dataset by a stable name while still being able to track exact snapshots used per experiment. You must register this in the Machine Learning workspace so runs can consume it reproducibly. Which approach should you use?

Accepted Answer

Register a data asset of type 'mltable' that points to the Parquet files with an MLTable definition, and let each registration create a new version.

Answer

Register a data asset of type 'uri_file' pointing directly to a single Parquet file so the schema is inferred automatically at runtime.

Answer

Create a new datastore for each dataset snapshot and reference the datastore name in the training job.

Answer

Store the Parquet path as an environment variable in the training environment and read it directly in the script.