Develop Direct Lake semantic models

This article describes design topics relevant to developing Direct Lake semantic models.

Create the model

You use the Fabric portal to create a Direct Lake semantic model in a workspace. It's a simple process that involves selecting which tables from a single lakehouse or warehouse to add to the semantic model.

You can then use the web modeling experience to further develop the semantic model. This experience allows you to create relationships between tables, create measures and calculation groups, mark date tables, and set properties for model and its objects (like column formats). You can also set up model row-level security (RLS) by defining roles and rules, and by adding members (Microsoft Entra user accounts or security groups) to those roles.

Alternatively, you can continue the development of your model by using an XMLA-compliant tool, like SQL Server Management Studio (SSMS) (version 19.1 or later) or open-source, community tools. For more information, see Model write support with the XMLA endpoint later in this article.

Model tables

Model tables are based on either a table or a view of the SQL analytics endpoint. However, avoid using views whenever possible. That's because queries to a model table based on a view will always fall back to DirectQuery mode, which might result in slower query performance.

Tables should include columns for filtering, grouping, sorting, and summarizing, in addition to columns that support model relationships. While unnecessary columns don't affect semantic model query performance (because they won't be loaded into memory), they result in a larger storage size in OneLake and require more compute resources to load and maintain.

To learn how to select which tables to include in your Direct Lake semantic model, see Edit tables for Direct Lake semantic models.

For more information about columns to include in your semantic model tables, see Understand storage for Direct Lake semantic models.

Enforce data-access rules

When you have requirements to deliver subsets of model data to different users, you can enforce data-access rules. You enforce rules by setting up object-level security (OLS) and/or row-level security (RLS) in the SQL analytics endpoint or in the semantic model.

Object-level security (OLS)

OLS involves restricting access to discover and query objects or columns. For example, you might use OLS to limit the users who can access the Salary column from the Employee table.

For a SQL analytics endpoint, you can set up OLS to control access to the endpoint objects, such as tables or views, and column-level security (CLS) to control access to endpoint table columns.

For a semantic model, you can set up OLS to control access to model tables or columns. You need to use open-source, community tools like Tabular Editor to set up OLS.

Row-level security (RLS)

RLS involves restricting access to subsets of data in tables. For example, you might use RLS to ensure that salespeople can only access sales data for customers in their sales region.

For a SQL analytics endpoint, you can set up RLS to control access to rows in an endpoint table.

For a semantic model, you can set up RLS to control access to rows in model tables. RLS can be set up in the web modeling experience or by using a third-party tool.

How queries are evaluated

The reason to develop Direct Lake semantic models is to achieve high performance queries over large volumes of data in OneLake. Therefore, you should strive to design a solution that maximizes the chances of in-memory querying.

The following steps approximate how queries are evaluated (and whether they fail). The benefits of Direct Lake storage mode are only possible when the fifth step is achieved.

1. If the query contains any table or column that's restricted by semantic model OLS, an error result is returned (report visual will fail to render).
2. If the query contains any column that's restricted by SQL analytics endpoint CLS (or the table is denied), an error result is returned (report visual will fail to render).
   1. If the cloud connection uses SSO (default), CLS is determined by the access level of the report consumer.
   2. If the cloud connection uses a fixed identity, CLS is determined by the access level of the fixed identity.
3. If the query contains any table in the SQL analytics endpoint that enforces RLS or a view is used, the query falls back to DirectQuery mode.
   1. If the cloud connection uses SSO (default), RLS is determined by the access level of the report consumer.
   2. If the cloud connection uses a fixed identity, RLS is determined by the access level of the fixed identity.
4. If the query exceeds the guardrails of the capacity, it falls back to DirectQuery mode.
5. Otherwise, the query is satisfied from the in-memory cache. Column data is loaded into memory as and when it's required.

Source item permissions

The account used to access data is one of the following.

• If the cloud connection uses SSO (default), it is the report consumer.
• If the cloud connection uses a fixed identity, it is the fixed identity.

The account must at least have Read and ReadData permissions on the source item (lakehouse or warehouse). Item permissions can be inherited from workspace roles or assigned explicitly for the item as described in this article.

Assuming this requirement is met, Fabric grants the necessary access to the semantic model to read the Delta tables and associated Parquet files (to load column data into memory) and data-access rules can be applied.

Data-access rule options

You can set up data-access rules in:

• The semantic model only.
• The SQL analytics endpoint only.
• In both the semantic model and the SQL analytics endpoint.

Rules in the semantic model

If you must enforce data-access rules, you should do so in the semantic model whenever viable. That's because RLS enforced by the semantic model is achieved by filtering the in-memory cache of data to achieve high performance queries.

It's also a suitable approach when report consumers aren't granted permission to query the lakehouse or warehouse.

In either case, it's strongly recommended that the cloud connection uses a fixed identity instead of SSO. SSO would imply that end users can access the SQL analytics endpoint directly and might therefore bypass security rules in the semantic model.

Rules in the SQL analytics endpoint

It's appropriate to enforce data-access rules in the SQL analytics endpoint when the semantic model cloud connection uses single sign-on (SSO). That's because the identity of the user is delegated to query the SQL analytics endpoint, ensuring that queries return only the data the user is allowed to access. It's also appropriate to enforce data-access rules at this level when users will query the SQL analytics endpoint directly for other workloads (for example, to create a Power BI paginated report, or export data).

Notably, however, a semantic model query will fall back to DirectQuery mode when it includes any table that enforces RLS in the SQL analytics endpoint. Consequently, the semantic model might never cache data into memory to achieve high performance queries.

Rules at both layers

Data-access rules can be enforced at both layers. However, this approach involves extra complexity and management overhead. In this case, it's strongly recommended that the cloud connection uses a fixed identity instead of SSO.

Comparison of data-access rule options

The following table compares data data-access setup options.

Recommended practices for enforcing data-access rules

Here are recommended practices related to enforcing data-access rules:

• If different users must be restricted to subsets of data, whenever viable, enforce RLS only at the semantic model layer. That way, users will benefit from high performance in-memory queries. In this case, it's strongly recommended that the cloud connection uses a fixed identity instead of SSO.
• If possible, avoid enforcing OLS and CLS at either layer because it results in errors in report visuals. Errors can lead to confusion or concern for users. For summarizable columns, consider creating measures that return BLANK in certain conditions instead of CLS (if possible).

Model write support with the XMLA endpoint

Direct Lake semantic models support write operations with the XMLA endpoint by using tools such as SSMS (19.1 or later), and open-source, community tools.

Before you can perform write operations, the XMLA read-write option must be enabled for the capacity. For more information, see Enable XMLA read-write.

Model write operations with the XMLA endpoint support:

• Customizing, merging, scripting, debugging, and testing Direct Lake model metadata.
• Source and version control, continuous integration and continuous deployment (CI/CD) with Azure DevOps and GitHub. For more information, see Content lifecycle management.
• Automation tasks like semantic model refresh, and applying changes to Direct Lake semantic models by using PowerShell and the REST APIs.

For more information, see Semantic model connectivity with the XMLA endpoint.

Direct Lake model metadata

When you connect to a Direct Lake semantic model with the XMLA endpoint, the metadata looks like that of any other model. However, Direct Lake models show the following differences:

• The compatibilityLevel property of the database object is 1604 (or higher).
• The mode property of Direct Lake partitions is set to directLake.
• Direct Lake partitions use shared expressions to define data sources. The expression points to the SQL analytics endpoint of the lakehouse or warehouse. Direct Lake uses the SQL analytics endpoint to discover schema and security information, but it loads the data directly from OneLake (unless it falls back to DirectQuery mode for any reason).

Post-publication tasks

After you publish a Direct Lake semantic model, you should complete some setup tasks. For more information, see Manage Direct Lake semantic models.

Unsupported features

The following model features aren't supported by Direct Lake semantic models:

• Calculated tables
• Calculated columns
• Hybrid tables
• User-defined aggregations
• Composite models, in that you can't combine Direct Lake storage mode tables with DirectQuery or Dual storage mode tables in the same model. However, it's technically possible to use Power BI Desktop to create a live connection to a Direct Lake semantic model and then extend it with new tables (using Import, DirectQuery, or Dual storage mode) or calculations. For more information, see Build a composite model on a semantic model.
• Columns based on SQL analytics endpoint columns that apply dynamic data masking.

Related content

• Direct Lake overview
• Manage Direct Lake semantic models
• Understand storage for Direct Lake semantic models
• Create a lakehouse for Direct Lake
• Edit tables for Direct Lake semantic models
• OneLake integration for semantic models