Lakehouse

Files and Delta tables in OneLake, written with Spark and read with SQL.

A Lakehouse is where most Fabric data ends up: raw files and clean tables, side by side, in OneLake. It is the default home for data your engineers load with Spark, and the most common source for Direct Lake reports.

Part of the Store layer of the ecosystem. Reviewed 13 November 2026 by Piyush Jain.

What Lakehouse is

A Lakehouse holds two things: a Files area for anything you land (CSV, JSON, Parquet) and a Tables area of Delta tables. Delta is an open format, so the same table can be read by Spark, by SQL and by a Power BI semantic model without copying it.

Every Lakehouse comes with a SQL analytics endpoint. It lets you query the tables with T-SQL and build views, but it is read-only: you load and change data with Spark notebooks, Dataflows Gen2 or pipelines, not with INSERT or UPDATE.

When to use it

Use it when

  • Your team writes Python or Spark, or you are landing files and semi-structured data.
  • You want a medallion layout: raw data in, cleaned tables out, then a gold layer for reporting.
  • The tables will feed a Direct Lake semantic model.

Think twice when

  • Your team works in T-SQL and needs stored procedures and multi-table transactions. That is the Warehouse.
  • You only need one report on one small source. Import mode on the source system is simpler.

The mistakes I see most

From opening other people's models. Each one is fixable.

  1. 1

    Loading everything into one layer

    Raw files, half-cleaned tables and reporting tables all in one place means nobody knows which table is the right one. Keep at least a raw layer and a reporting layer, even in a small setup.

  2. 2

    Thousands of tiny files

    Frequent small loads leave a Delta table split into many small Parquet files. Reads slow down and Direct Lake has more work to do. Schedule table maintenance (OPTIMIZE and VACUUM) as part of the load, not as an afterthought.

  3. 3

    Building report logic in endpoint views

    Views on the SQL analytics endpoint are handy, but a Direct Lake model that reads a view falls back to DirectQuery and loses the speed you moved to Fabric for. Materialise reporting tables as Delta tables instead.

Questions people ask

What is the difference between a Lakehouse and a Warehouse in Fabric?
Both store Delta tables in OneLake. A Lakehouse is written with Spark and read with SQL through a read-only endpoint. A Warehouse is written and read with T-SQL, with full DML and multi-table transactions. Pick the one that matches the skills of the team that will load the data.
Can I write to a Lakehouse with SQL?
Not through the SQL analytics endpoint, which is read-only. Use a notebook, a Dataflow Gen2 or a pipeline to load and change data, or use a Warehouse if you need T-SQL writes.
Does Power BI read a Lakehouse directly?
Yes. A semantic model in Direct Lake mode reads the Delta tables in OneLake without an import refresh. Import and DirectQuery through the SQL analytics endpoint also work.

Sources: Microsoft Learn: What is a lakehouse in Microsoft Fabric?; Microsoft Learn: Direct Lake overview. Opinions are my own.