Skip to main content
direct lakedirect lake modedirect lake vs importdirect lake on onelakedirectquery fallbackpower bi fabric

What Is Direct Lake? Power BI Storage Mode in Fabric

Direct Lake is a storage mode for Power BI semantic models in Microsoft Fabric. A Direct Lake model reads its tables straight from Delta tables in OneLake and loads the columns a query needs into memory on demand, so it answers queries at in-memory speed without the scheduled data copy that Import mode needs. Microsoft describes it as "a Power BI semantic model table storage mode option available in Microsoft Fabric," and it exists only on Fabric capacity.

This guide explains how Direct Lake works, how it compares with Import and DirectQuery, the difference between its two variants, and the guardrails and limitations that decide whether it fits your models.

TL;DR

Direct Lake lets a Power BI semantic model query Delta tables in OneLake directly. A refresh copies only metadata (called framing) and takes seconds, and columns are loaded into memory the first time a query asks for them (transcoding). There are two variants: Direct Lake on OneLake, which Microsoft recommends for new models, never falls back to DirectQuery and supports composite models; Direct Lake on SQL goes through a SQL analytics endpoint and can fall back to DirectQuery when a table exceeds capacity guardrails or SQL-based security is in play. Direct Lake needs a Fabric capacity (F-SKU), and guardrails such as rows per table scale with the SKU, from 300 million rows on an F2 to 1.5 billion on an F64.

What Is Direct Lake?

Every Power BI semantic model stores each table in a storage mode. Until Fabric, there were two main options. Import copies the data into the model's in-memory VertiPaq engine, which is fast to query but must be refreshed on a schedule and duplicates the data. DirectQuery leaves the data in the source and sends a query there every time a visual renders, which keeps data current but is slower and loads the source system.

Direct Lake is a third option built for data that already sits in OneLake as Delta tables, for example in a Fabric lakehouse or warehouse. Queries are still processed by the VertiPaq engine, as with Import, but the data is not copied into the model by a scheduled refresh. Instead, the model points at the Delta files and loads the columns it needs when it needs them. For large tables that change often, this removes the trade-off between fast reports and fresh data that used to define Power BI model design.

How Direct Lake Works

Microsoft's documentation on how Direct Lake works describes three mechanisms.

Framing. A Direct Lake refresh does not copy data. It records which version of each Delta table the model should use, a metadata operation Microsoft calls framing, which "can take a few seconds to complete." Framing "gives model owners point-in-time control over what data loads into the semantic model": queries see the state of the Delta table at the most recent successful framing, even if new data has landed since. A setting called "Keep your Direct Lake data up to date," enabled by default, frames the affected tables automatically when the data changes.

Transcoding. "Direct Lake semantic models load data from OneLake only when a query requests a column for the first time. This on-demand data loading process is known as transcoding." Columns nobody queries are never loaded, which keeps memory use proportional to what reports actually use.

Eviction. Loaded columns stay in memory until there is a reason to remove them: a refresh after the source data changed, a column that no query has used for some time, or memory pressure on the capacity. When a model needs more memory than its SKU allows, columns are paged in and out, which costs query performance rather than failing outright.

HOW A DIRECT LAKE MODEL GETS ITS DATA OneLakeDelta tables in alakehouse or warehouseParquet files Semantic modelVertiPaq in-memory engineFraming: table versionColumns loaded on demand Power BI reportsDAX queries from visualsin-memory speed transcodequery A refresh copies metadata only. Data is read from OneLake the first time a query needs a column.
Click to enlarge

Direct Lake vs Import vs DirectQuery

DimensionImportDirectQueryDirect Lake
Where data is read fromA copy inside the semantic modelThe source system, at query timeDelta tables in OneLake, loaded into memory on demand
Query engineVertiPaq (in memory)The source databaseVertiPaq (in memory)
RefreshFull or incremental data reload on a scheduleNone neededFraming: metadata only, seconds
Data freshnessAs of the last refreshLiveAs of the last framing, which can run automatically
Query speedFastDepends on the sourceFast once columns are loaded
SourcesAlmost anySupported databasesDelta tables in OneLake only
LicensingWorks without a capacityWorks without a capacityFabric capacity required

Import still wins for small and medium models from sources outside OneLake, and DirectQuery for live operational data that cannot be copied. Direct Lake is the better fit when the data is already curated in OneLake, the tables are large, and reports must reflect new loads quickly. Composite models can mix modes, and Direct Lake on OneLake tables can be combined with Import tables in one model.

Direct Lake on OneLake vs Direct Lake on SQL

Direct Lake comes in two variants, and they behave differently in ways that matter for design and security. Microsoft's Direct Lake overview compares them, and its how-it-works page says of Direct Lake on OneLake: "For new semantic models, it's the recommended Direct Lake option."

DimensionDirect Lake on OneLakeDirect Lake on SQL
Data sourcesOne or more Fabric data sourcesA single Fabric data source, via its SQL analytics endpoint
DirectQuery fallbackNever: runs in DirectLakeOnly modeCan fall back to DirectQuery through the SQL endpoint
When a guardrail is exceededRefresh fails and the model cannot be queriedQueries fall back to DirectQuery; refresh succeeds with a warning
Composite modelsSupportedNot supported
Calculated columnsUser-context only, in previewNot supported
SQL endpoint row-level securityNot applied; users need access to the files in OneLakeCauses fallback to DirectQuery
Recommended for new modelsYesNo

The practical effect is that Direct Lake on OneLake gives predictable performance but fails hard when a table outgrows the capacity, while Direct Lake on SQL degrades gracefully to slower DirectQuery. Teams moving from the SQL variant should check table sizes against the guardrails first.

DirectQuery Fallback and Guardrails

Guardrails are per-SKU resource limits. In the SQL variant, a query that exceeds them falls back to DirectQuery; in the OneLake variant, it fails. Most are evaluated per query; model size is evaluated for the whole model. A sample of the published table:

Fabric SKUParquet files per tableRow groups per tableRows per tableMax model size on diskMax memory
F21,0001,000300 million10 GB3 GB
F64 (and P1)5,0005,0001.5 billionUnlimited25 GB
F204810,00010,00024 billionUnlimited400 GB

Source: Direct Lake overview, checked 16 November 2026. "Max memory" is not a guardrail: exceeding it causes paging rather than fallback.

For Direct Lake on SQL, a model property called DirectLakeBehavior controls fallback. Automatic (the default) falls back to DirectQuery silently, DirectLakeOnly returns an error instead, and DirectQueryOnly always uses DirectQuery. The DAX function TABLETRAITS() shows, per table, whether and why a query would fall back. Because the file and row-group counts matter as much as row counts, Delta table maintenance (compacting small files) is part of keeping a Direct Lake model inside its guardrails. Framing itself can fail when a table exceeds them, for example with more than 10,000 Parquet files.

Direct Lake Requirements and Licensing

Direct Lake has firmer prerequisites than Import or DirectQuery.

  • Fabric capacity. "Direct Lake semantic models require a Fabric capacity license." Licensing for both variants is Fabric capacity (F-SKU) only, so Direct Lake is not available on Pro or Premium Per User alone. See Microsoft Fabric pricing for the capacity ladder and Microsoft Fabric vs Power BI for when capacity makes sense.
  • Delta tables in OneLake. The source tables must be Delta tables in a Fabric lakehouse, warehouse, mirrored database or other OneLake item. Data elsewhere can be brought in with OneLake shortcuts or mirroring.
  • Cloud connections only. Direct Lake "supports only cloud connections and cannot operate through any gateway."
  • Per-user licenses still apply for viewers. On capacities smaller than F64, every viewer of Power BI content still needs a Pro or PPU license.

Direct Lake Limitations

Per Microsoft's documentation, these are the limitations most likely to affect a design:

  • Hybrid tables and custom table partitions are not supported in either variant.
  • Complex Delta column types, and binary and GUID semantic types, are not supported.
  • String values are limited to 32,764 Unicode characters.
  • Non-numeric floating-point values such as NaN are not supported.
  • Calculated columns are available only in the OneLake variant, only with user context, and only in preview. Most calculation logic belongs upstream in the lakehouse or warehouse, or in DAX measures.
  • Composite models are supported only in the OneLake variant.

None of these limits are unusual for a lakehouse-first design, but they reward modeling discipline: clean, typed, compacted Delta tables in OneLake and a semantic model that only adds relationships and measures.

Direct Lake Security and Governance

Direct Lake changes where security is enforced. Semantic-model row-level and object-level security work in both variants, and Microsoft strongly recommends a fixed-identity cloud connection when row-level security is defined in the model. In the OneLake variant, row-level security defined in the SQL analytics endpoint is not applied, and "Direct Lake on OneLake requires the user has access to the files in OneLake," so OneLake security roles become part of report access design. In the SQL variant, SQL endpoint row-level security forces fallback to DirectQuery.

Direct Lake also removes the natural checkpoint that Import refresh used to provide. A change to a lakehouse table now reaches reports within minutes, so knowing which semantic models and reports depend on which tables, and who owns them, matters more than before. Dawiso's Microsoft Fabric connector ingests lakehouses with their tables and columns, warehouses, semantic models with their columns, and reports into Dawiso's data catalog, where owners, business glossary definitions and descriptions sit next to the technical objects. That gives the people changing a lakehouse table and the people reading the report a shared reference for what the data means. For the broader picture of Power BI in Fabric, see our Power BI Fabric integration guide and our guide to Power BI semantic models.

Sources

See it in action

Data & Analytics Catalog

Create a unified view of your data assets and gain insights faster with automated data discovery.

A cookie a day keeps bad UX away.

We use cookies to personalize content, ads and to analyze our traffic. We also share information about your use of our site with our advertising and analytics partners who may combine it with other information that you've provided to them or that they've collected from your use of their services. By clicking "Accept All", you allow us to use cookies for analytics and ads via Google Tag Manager. You can also customize cookies.

Customize Consent Preferences

We use cookies to personalize content, ads and to analyze our traffic. We also share information about your use of our site with our advertising and analytics partners. Privacy Policy

Necessary cookies allow core website functionality such as user login and account management. The website cannot be used properly without strictly necessary cookies.

Functionality cookies are used to remember visitor information on the website, eg. language, timezone, enhanced content.

Analytics cookies are used to see how visitors use the website, eg. analytics cookies. Those cookies cannot be used to directly identify a certain visitor.

We use Microsoft Clarity to see how you use our website (including heatmaps and session replays) so we can improve it. By using our site, you agree that we and Microsoft can collect and use this data. See our Privacy Policy for details.

Advertisement cookies are used to identify visitors between different websites, eg. content partners, banner networks. Those cookies may be used by companies to build a profile of visitor interests or show relevant ads on other websites.