Fabric Lakehouse

Fabric Lakehouse

Core Concepts

Lakehouse in Microsoft Fabric combines the flexibility of a data lake with the management of a data warehouse:

• Unified storage in OneLake for structured and unstructured data
• Delta Lake format for ACID transactions, versioning, and time travel
• SQL analytics endpoint for T-SQL queries (auto-generated, read-only)
• Default semantic model for Power BI integration
• Support for CSV, Parquet (Spark-only querying for non-Delta formats)

Key Components

Schemas

When creating a Lakehouse, users can enable schemas to organize tables:

• Schemas are folders under Tables/
• Default schema is dbo (cannot be deleted or renamed)
• Can reference schemas in other Lakehouses via Schema Shortcuts

Security

Workspace Roles (Control Plane)

OneLake Security (Data Plane)

• Based on Microsoft Entra ID and RBAC
• Supports column-level and row-level security on tables
• Data access controlled through OneLake permissions

Shortcuts

Virtual links to data without copying:

Create Shortcut (REST)

POST /v1/workspaces/<wsId>/items/<lakehouseId>/shortcuts
{
  "path": "Tables/external_data",
  "name": "my_shortcut",
  "target": {
    "oneLake": {
      "workspaceId": "<sourceWsId>",
      "itemId": "<sourceItemId>",
      "path": "Tables/source_table"
    }
  }
}

Table Optimization

V-Order

Write optimization that applies sorting, encoding, and compression for fast reads. Enabled by default for Fabric Spark.

df.write.format("delta").option("vorder", "true").save(path)

OPTIMIZE and VACUUM

-- Compact small files
OPTIMIZE lakehouse.schema.table_name

-- Remove old files (default 7-day retention)
VACUUM lakehouse.schema.table_name

Z-Order

Co-locate related data for faster filter queries:

OPTIMIZE lakehouse.schema.table_name ZORDER BY (column1, column2)

PySpark Patterns

Read Delta Table

df = spark.read.format("delta").load("Tables/my_table")
# or
df = spark.sql("SELECT * FROM lakehouse.dbo.my_table")

Write Delta Table

df.write.format("delta") \
    .mode("overwrite") \
    .option("overwriteSchema", "true") \
    .save("Tables/my_table")

Incremental Load (Merge)

from delta.tables import DeltaTable

target = DeltaTable.forPath(spark, "Tables/my_table")
target.alias("t").merge(
    source_df.alias("s"),
    "t.id = s.id"
).whenMatchedUpdateAll() \
 .whenNotMatchedInsertAll() \
 .execute()

Notebook Utilities

Microsoft Fabric renamed MSSparkUtils to NotebookUtils. Existing mssparkutils code remains backward-compatible, but new code should prefer notebookutils for continued support and access to newer modules.

# List files
notebookutils.fs.ls("Files/raw/")

# Copy files
notebookutils.fs.cp("Files/source/", "Files/dest/", recurse=True)

# Get secret from Key Vault
secret = notebookutils.credentials.getSecret("https://my-kv.vault.azure.net/", "secret-name")

# Run another notebook
notebookutils.notebook.run("other_notebook", timeout_seconds=600, arguments={"param1": "value1"})

Medallion Architecture

Each layer is a separate Lakehouse or schema within a Lakehouse.

Must

• Use Delta format for all managed tables
• Enable V-Order for read-heavy workloads
• Partition large tables by date or high-cardinality columns
• Run OPTIMIZE regularly on frequently queried tables
• Use shortcuts instead of copying data across workspaces

Avoid

• Storing non-Delta tabular data if SQL endpoint access is needed
• Deep folder nesting in Files/ (keep hierarchy shallow)
• Small file accumulation without OPTIMIZE
• Hardcoding OneLake paths — use workspace/item variables