Direct Lake semantic models

Last checked against Microsoft documentation: 14 November 2026

Direct Lake is a table storage mode in Microsoft Fabric. The semantic model reads Delta tables from OneLake straight into memory, so you get import-style speed without copying the data at every refresh. Queries are run by the same VertiPaq engine that import mode uses.

How it compares

ImportDirectQueryDirect Lake
Where the data livesCopied into the modelStays at the sourceStays in OneLake as Delta/Parquet files
What a refresh doesCopies the dataNothing to refreshCopies only metadata ("framing"), usually in seconds
Who runs the queryVertiPaqThe source databaseVertiPaq
LicenceAny Fabric or Power BI licenceAny Fabric or Power BI licenceFabric capacity only

Only the columns a query needs are loaded, so the data you analyse can be larger than the capacity's memory limit.

Two ways to build it

Microsoft recommends Direct Lake on OneLake for new models.

Direct Lake on OneLakeDirect Lake on SQL endpoint
Data sourcesOne or more Fabric items with Delta tables, including shortcutsOne supported Fabric item with a SQL analytics endpoint
Falls back to DirectQueryNoYes, if fallback is enabled
SQL viewsNot supported (use a lakehouse materialized view)Supported, but they fall back to DirectQuery
Mix with Import tablesYesNo
Calculated columnsUser context only (preview)Not supported
Calculated tablesYes (preview)Not supported, except calculation groups, what-if and field parameters

How it works: three ideas

1. Transcoding is loading columns on demand. The first query that needs a column triggers its load from OneLake. That includes columns used by relationships and measures. A loaded column is called resident until it is evicted. Eviction happens after a framing change, after a period of no use, or under memory pressure.

2. Framing is the "refresh" of Direct Lake. It reads the latest Delta table metadata and points the model at the newest Parquet files. It sets a point-in-time baseline: queries see the tables as they were at the last successful framing. Only the affected column segments are dropped, and dictionaries are usually kept.

Framing runs in four ways: automatically, manually, on a schedule, or by code (Power BI REST API or TOM). The setting Keep your Direct Lake data up to date is on by default. Switch it off when an ETL job must finish before new data is shown. If a non-recoverable error happens, Power BI suspends automatic updates until a later manual refresh succeeds.

3. Fallback exists only for Direct Lake on SQL endpoint. A query stays in Direct Lake only if no referenced table has SQL row-level security, dynamic data masking or object-level security, none is an unmaterialized SQL view, no table exceeds the capacity guardrails, and the model was framed after the Delta tables last changed.

Setting DirectLakeBehaviorWhat happensUse it for
Automatic (default)Quietly falls back to DirectQueryProduction
DirectLakeOnlyThe query fails with an errorDevelopment, to find problems
DirectQueryOnlyAlways uses DirectQueryDevelopment, to measure fallback speed

Direct Lake on OneLake always behaves like DirectLakeOnly.

Memory states: cold, semiwarm, warm, hot

StateWhat it means
ColdNothing is in memory. Every needed column loads from Delta
SemiwarmOnly segments from removed row groups were dropped. Only new or changed data loads
WarmThe needed columns are fully in memory
HotColumns are in memory and query caches are filled

Speed improves from cold to hot. Your goal is to reach warm quickly and avoid going back to cold.

Step by step: build and tune

  1. Use your Gold layer. Point Direct Lake at curated serving tables. Avoid raw Bronze tables unless you designed them for this.
  2. Create the model. In the browser, open a lakehouse and choose New semantic model. You can also create and live-edit it in Power BI Desktop. Choose Direct Lake on OneLake.
  3. Check the guardrails for your capacity (table below).
  4. Prepare the Delta tables. For tables mainly read by Direct Lake, turn on V-Order, aim for row groups of 1 to 16 million rows, and keep the Parquet file count low. Leave Fabric's adaptive file size (128 MB to 1 GB) and auto compaction on, and keep deletion vectors enabled.
  5. Choose an update pattern (table below). Prefer appends.
  6. Decide on framing. Keep automatic updates on for frequently changing tables, or turn them off and frame after your ETL finishes.
  7. Use integer keys instead of text where you can, and reduce columns with very many distinct values.
  8. Turn off DirectQuery fallback where possible to avoid mixed query plans.
  9. Test it with the two queries in the next section.

Capacity guardrails (selected sizes)

Fabric SKUParquet files per tableRow groups per tableRows per table (millions)Max memory (GB)
F21,0001,0003003
F645,0005,0001,50025
F1285,0005,0003,00050
F51210,00010,00012,000200
F1024 and above10,00010,00024,000400

Max memory is not a guardrail: going over it causes paging and slower queries. On F2 the model size on disk is also capped at 10 GB. The model-size limit applies to the whole model, while the other limits are checked per query. When a limit is exceeded, Direct Lake on OneLake fails to refresh until you optimize the tables, and Direct Lake on SQL endpoint falls back to DirectQuery.

Update patterns: what each does to performance

Parquet files cannot be changed in place, so an UPDATE is really a delete plus an append.

PatternEffect on Direct Lake
Overwrite the tableErases the Delta log, so every column is loaded again from cold
Batch load, no partitions, rolling-window deletesDeletes touch most files. Removing 1 day of 30 can affect about 95% of files
Batch load with partitionsDeletes stay inside the affected partitions. Use a partition column with fewer than 100 to 200 distinct values
Append onlyBest case. Existing files stay untouched and incremental framing works
StreamingCreates many small files and row groups. Run OPTIMIZE daily or more often

Maintenance rules

  • OPTIMIZE merges small files. It clears in-memory segments, so expect a short cold period. Run it on a calm schedule, such as weekends, unless small files pile up quickly. It also purges deletion vectors on files where more than 5% of records are covered.
  • VACUUM removes old files and does not affect speed. A framed model points to one specific Delta version, so keep that version until you reframe. Otherwise queries fail because files are missing.
  • Do not shorten the VACUUM retention period without checking time-travel and concurrent readers.

Two checks you can run in DAX query view

Find out why a table fell back (on the SQL endpoint variant):

DAX
EVALUATE TABLETRAITS()

Read the DirectLakeFallbackInfo column. None means the table is in Direct Lake mode.

Check whether framing kept your data in memory:

DAX
EVALUATE INFO.STORAGETABLECOLUMNSEGMENTS()

Query until the model is warm, update the Delta table, refresh the model, then run this query. If segments are no longer resident, the model reloaded from cold, which usually points to a destructive update pattern.

Security and connections

  • Object-level security works in both variants. Semantic model row-level security also works in both. Microsoft strongly recommends a fixed-identity cloud connection with it.
  • With Direct Lake on OneLake, SQL-based RLS is not applied. Users need access to the OneLake files.
  • With Direct Lake on SQL endpoint, SQL RLS is enforced by falling back to DirectQuery. If fallback is off, the query fails.
  • Only cloud connections work. Gateways, on-premises or VNet, are not supported. Service principal profiles are not supported. Embedding needs a V2 embed token.

Watch-outs

  • Complex Delta column types, Binary and GUID types, and NaN values are not supported. Convert them first.
  • Relationship columns must have matching data types, and the one-side column must be unique.
  • Auto date/time for relationships is not supported. Mark your own date table.
  • Text values are limited to 32,764 characters.
  • No hybrid tables, no table partitions in the model (partition the Delta table instead), and no user-defined aggregations.
  • Models cannot be created in My workspace, or in a different region from the lakehouse or warehouse.
  • Deployment pipeline data source rules are not supported for Direct Lake on OneLake. Use M parameters and parameter rules.
  • XMLA tools need compatibility level 1604 or higher. Tables created through XMLA start unprocessed, so run a refresh.
  • Direct Lake does not use Delta or Parquet statistics to skip files or row groups while loading.
  • A model can reload cold on another Fabric node if compute or memory demand changes unexpectedly.

Official sources