Azure ML Storage Accounts vs Datastores vs Data Assets (URI File, URI Folder, MLTable) – When should each be used?

• Storage account: the physical place where data lives.
• Datastore: an Azure Machine Learning reference to an existing Azure storage account. It gives a common way to access Blob, Files, and ADLS storage, and can use either credential-based or identity-based authentication.
• Data asset: a named reference to a data location plus metadata. It behaves like a bookmark to frequently used data, without copying the data or adding storage cost.

For Azure ML data types:

• uri_file: use when the input is a single file of any format.
• uri_folder: use when the input is a folder of files.
• mltable: use when the data is a data table, especially when schema can change, when a subset of large tabular data is needed, for AutoML with tables, or when data is spread across multiple storage locations.

Practical differences:

1. uri_file
   • Best for one file.
   • Example scenarios from the docs: a single CSV, a single parquet file, or any one file.
2. uri_folder
   • Best for a folder of files.
   • Example scenarios: a folder of parquet/CSV files, or unstructured data such as images, text, and audio stored in a folder.
3. mltable
   • Best when the data should be treated as a table abstraction rather than just files.
   • Recommended for complex or changing schema, large tabular subsets, AutoML with tables, and data spread across multiple storage locations.
   • It can also be used for unstructured data such as images, text, and audio when that data is spread across multiple storage locations.
   • For CSV files with embedded newlines, mltable is the recommended choice because read_delimited supports support_multi_line.

For the formats listed, the supported guidance is:

• CSV
  • Single CSV file: uri_file
  • Folder of CSV files: uri_folder
  • CSV with embedded newlines, or when schema handling is important: mltable
• Parquet
  • Single parquet file: uri_file
  • Folder of parquet files: uri_folder
  • If table semantics, schema evolution, or subsetting are needed: mltable
• Images
  • Folder of images: uri_folder
  • If image data is spread across multiple storage locations and should be handled as a table abstraction: mltable
• Text files
  • Single text file: uri_file
  • Folder of text files: uri_folder
  • If files are spread across multiple storage locations and need table-style handling: mltable
• Excel (.xlsx)
• JSON
  • The provided material does not give format-specific recommendations for these beyond the general rule that uri_file is for a single file and uri_folder is for a folder.

On whether MLTable is only for tabular datasets:

• No. It is a table abstraction, but the documented scenarios also include unstructured data such as images, text, and audio, especially when that data is distributed across multiple storage locations.

When to register a data asset instead of using a datastore path directly:

• Register a data asset when the same data is used frequently and a friendly name is preferable to a long URI.
• Use a data asset when metadata alongside the reference is useful.
• A data asset is useful for repeatable project organization because it creates a named reference to the source location.

Benefits of data assets over direct storage references that are explicitly supported:

• Usability: avoids remembering long URIs.
• Metadata: stores a copy of metadata with the reference.
• No extra storage cost: the data stays in its original location.
• No risk to source integrity from registration: creating the asset does not move or duplicate the data.

For direct datastore access:

• Datastores are useful when a team needs a common access layer to storage.
• Datastore URIs can be used directly, including through fsspec, and can be read like filesystems in Python tools such as Pandas.
• Datastores also help avoid exposing secrets in scripts because authentication can be handled through the datastore.

A practical production-oriented strategy supported by the docs is:

1. Keep the actual data in Azure storage.
2. Register that storage as a datastore for secure and reusable access.
3. Use data assets for important datasets that should have stable, friendly names.
4. Choose the asset type by shape of data:
   • one file -> uri_file
   • one folder -> uri_folder
   • table abstraction / evolving schema / large tabular subsets / multi-location data -> mltable
5. Prefer URIs for most scenarios unless table-specific capabilities are needed; the glossary explicitly encourages starting with uri_folder and uri_file if not using AutoML.

For jobs, Azure ML maps URIs to the compute filesystem, and data can be consumed by mount or download modes. That makes URI-based inputs practical for both direct references and registered assets.

References:

• Data concepts in Azure Machine Learning
• Create and manage data assets
• Access data from Azure cloud storage during interactive development
• Azure Machine Learning glossary