Fabric Lakehouse Tutorial — Complete 2026 Guide

What Creates When You Create a Lakehouse

When you create a Fabric Lakehouse, the platform automatically provisions three connected items. Understanding this prevents the most common beginner confusion about why tables appear in some interfaces but not others.

June 2026 Updates — What Changed

The June 2026 Fabric release (published June 2, 2026) added two significant Lakehouse-specific features and several OneLake platform improvements that affect cost management and cross-workspace engineering patterns.

Architecture & Medallion Design

Fabric Lakehouse architecture centres on OneLake — a single, tenant-wide logical data lake backed by Azure Data Lake Storage Gen2 with a hierarchical namespace. Every workspace appears as a container, and every Lakehouse appears as a folder inside that container. Spark, T-SQL, Power BI, and pipelines all read the same underlying Parquet files without copying data between engines.

Core Components

Medallion Architecture in Fabric Lakehouse

Microsoft recommends organising Lakehouse data into Bronze, Silver, and Gold layers inside a single Lakehouse using Lakehouse schemas (available when Lakehouse schemas are enabled) or across multiple Lakehouses connected by OneLake shortcuts and Materialized Lake Views.

Materialized Lake Views — GA & June 2026 Updates

Materialized Lake Views reached General Availability in March 2026. They let you define transformation logic in SQL or PySpark — Fabric materialises the result as a Delta table and manages incremental refresh, dependency tracking, and retry automatically. The June 2026 release added cross-workspace extended lineage, changing how multi-Lakehouse Medallion orchestration works.

What MLVs Replace

Before MLVs, a Silver-to-Gold transformation required a Spark notebook triggered by a pipeline, with manual dependency management between tables. An engineer had to know that Gold table A depends on Silver table B which depends on Bronze table C, and schedule three separate pipelines in the right order. MLVs declare these dependencies as SQL or PySpark, and Fabric resolves the execution order automatically — including across Lakehouses as of June 2026.

Key Capabilities (Post-GA, March 2026)

Gold Sales Summary — Materialized Lake View Definition

-- Create a Gold-layer summary MLV
-- Source tables live in Silver schema (same or other Lakehouse via shortcut)
CREATE MATERIALIZED LAKE VIEW gold.sales_summary
AS
SELECT
    c.region,
    p.category,
    SUM(f.sales_amount)            AS total_sales,
    COUNT(*)                       AS txn_count,
    COUNT(DISTINCT f.order_id)    AS unique_orders
FROM   silver.fact_sales     f
JOIN   silver.dim_customer   c  ON f.customer_id = c.customer_id
JOIN   silver.dim_product    p  ON f.product_id  = p.product_id
GROUP BY c.region, p.category;

-- After creation: appears as a Delta table in the Lakehouse explorer
-- Queryable via SQL endpoint and Power BI Direct Lake
-- Refresh schedule: Lakehouse → Materialized lake views → Schedules

Loading Data into Fabric Lakehouse

Fabric supports multiple ingestion paths into Lakehouse, all landing data in OneLake. The right method depends on source type, latency requirement, and whether the team prefers code-first or low-code tooling.

Data Pipelines — Copy Activity and Copy Job

Data Pipelines connect to 90+ sources and land data into Lakehouse tables or Files. Copy Activity gives full control over schema mapping, partition options, and fault tolerance. Copy Job Activity (GA April 2026) is the no-code equivalent with automatic schema detection, bulk/incremental/CDC in one item, and SCD Type 2 support for Fabric Warehouse. CDC for SQL Server, Azure SQL, and Azure SQL MI is now GA; auto-partition support for Oracle, Lakehouse, and SAP HANA is in Preview.

# Read raw files from ADLS landing zone
df = (spark.read
      .format("parquet")
      .load("abfss://landing@storage.dfs.core.windows.net/raw/sales/"))

# Write as managed Delta table in Bronze schema
(df.write
   .format("delta")
   .mode("overwrite")
   .option("overwriteSchema", "true")
   .saveAsTable("bronze.sales_raw"))

Dataflow Gen2 — Low-Code Ingestion

Dataflow Gen2 uses Power Query with 200+ connectors to load and transform data into Lakehouse tables — no Spark knowledge required. The June 2026 release added Mapping Data Flow transformations inside Dataflow Gen2, bringing ADF Mapping Data Flow logic into the Fabric low-code experience. Output directly to Lakehouse managed tables, queryable immediately via the SQL endpoint and Power BI.

Spark Notebooks — Complex ETL and CDC

Notebooks are the right tool for complex ingestion: CDC with Delta MERGE, custom schema enforcement, data quality gates, feature engineering, and multi-source joins. Use mssparkutils.runtime.getParameterValue() to receive pipeline parameters so notebooks are parameterised and rerunnable rather than hardcoded.

from delta.tables import DeltaTable
from pyspark.sql.functions import current_timestamp, lit

RUN_DATE = mssparkutils.runtime.getParameterValue("run_date", "2026-06-06")

# Read incremental source data
df_inc = (spark.read.format("delta")
          .table("bronze.sales_raw")
          .filter(col("load_date") == lit(RUN_DATE))
          .dropDuplicates(["order_id"]))

# MERGE to Silver table — safe to rerun for the same run_date
target = DeltaTable.forName(spark, "silver.fact_sales")
(target.alias("tgt")
       .merge(df_inc.alias("src"), "tgt.order_id = src.order_id")
       .whenMatchedUpdateAll()
       .whenNotMatchedInsertAll()
       .execute())

Eventstream — Real-Time Ingestion

Fabric Eventstream ingests streaming data from Azure Event Hubs, IoT Hub, Kafka, MQTT, and CDC sources. Route streams to Lakehouse Delta tables, Eventhouse KQL databases, or both simultaneously. June 2026 added GA support for Apache Kafka and Azure Service Bus connectors, pagination support for the HTTP connector (Preview), and Eventstream observability through Workspace Monitoring (Preview).

Mirroring and OneLake Shortcuts

Data Mirroring keeps operational databases (Azure SQL DB, Cosmos DB, Snowflake, Databricks Unity Catalog) in sync as Delta tables in OneLake without custom ETL — near-real-time replication with zero code. OneLake shortcuts virtualise data from ADLS Gen2, S3, or other Lakehouses as folders inside your Lakehouse without copying it. June 2026 added network security controls for Mirroring — workspace-level outbound access protection for mirrored connections.

SQL Endpoint, Direct Lake & Power BI

The SQL analytics endpoint is auto-generated at Lakehouse creation and exposes every managed Delta table as a read-only T-SQL relation. Connect with Power BI Desktop, Azure Data Studio, or any T-SQL client using Entra ID authentication against the Fabric connection endpoint.

SELECT
    product_name,
    SUM(revenue)    AS total_revenue,
    COUNT(order_id) AS order_count
FROM  gold.sales_summary
GROUP BY product_name
ORDER BY total_revenue DESC;

June 2026 — SQL Endpoint Updates

The June 2026 release added several SQL analytics endpoint improvements: CI/CD Support for SQL analytics endpoint (Preview) — version-control and promote SQL endpoint objects across environments; Enhanced metadata sync (Preview) — improved consistency between Lakehouse table changes and SQL endpoint schema; Time Travel via SQL analytics endpoint (Preview) — query historical Delta table versions using SQL syntax; and Configurable Retention (Preview) for controlling Delta history depth.

Power BI Direct Lake Mode

Direct Lake allows the Power BI Analysis Services engine to read Delta Parquet files directly from OneLake memory — no import job, no scheduled refresh window. Reports reflect data as soon as the upstream ETL or MLV refresh writes new files to the Gold Delta table.

Security & Governance

Fabric Lakehouse security operates at three layers: workspace RBAC, item-level permissions, and data-plane controls. All identity is anchored to Microsoft Entra ID — no separate credential store.

OneLake Shortcuts — Permission Requirements

When creating shortcuts to external ADLS Gen2, S3, or cross-workspace Lakehouses, the identity accessing the shortcut (your Entra ID in interactive sessions, the workspace identity or service principal in pipeline runs) must have at least Storage Blob Data Reader on the target ADLS account, or equivalent permissions on the target Lakehouse workspace. Shortcut authorization failures are almost always a permission mismatch between the accessing identity and the target storage. See the OneLake shortcut authorization fix guide for the full diagnostic checklist.

Governance and Lineage

Microsoft Purview captures lineage from sources through Lakehouse ingestion pipelines, MLVs, SQL endpoint views, semantic models, and Power BI reports. As of 2026, cross-workspace lineage for non-Power BI Fabric assets has a known limitation in Purview — MLV extended lineage (June 2026) addresses the orchestration side of this by making refresh dependency order visible inside Fabric itself, while the Purview cross-workspace lineage gap remains on the roadmap.

Optimization, Maintenance & Cost Control

Performance and cost in a Fabric Lakehouse depend on Delta file layout, partitioning strategy, query patterns, and OneLake storage tier decisions. The June 2026 storage tiers feature adds a new cost lever that did not exist previously.

OneLake Storage Tiers — Cost Strategy

All OneLake data defaults to the Hot tier. The three tiers have different cost trade-offs: Hot has the lowest transaction cost and is correct for Delta tables actively queried by Spark, SQL endpoint, or Direct Lake. Cool has lower storage cost but higher transaction cost — suitable for historical partitions accessed monthly or less. Cold has the lowest storage cost with the highest transaction cost and retrieval charges — suitable for compliance archives accessed less than once per year.

# Example lifecycle rule (configured in Fabric UI or via API, not T-SQL)
# Rule: Move files not accessed in 90 days to Cool tier
Rule name:  move_old_bronze_to_cool
Scope:      /workspaces/{workspace_id}/lakehouses/{lakehouse_id}/Tables/bronze.*/
Action:     ChangeTier → Cool
Condition:  LastAccessedDaysAgo >= 90

# Rule: Move files not accessed in 365 days to Cold tier
Rule name:  move_archive_to_cold
Scope:      /workspaces/{workspace_id}/lakehouses/{lakehouse_id}/Tables/bronze.*/
Action:     ChangeTier → Cold
Condition:  LastAccessedDaysAgo >= 365

# NOTE: Delta transaction log files should remain Hot —
# exclude _delta_log/ paths from cooling rules

Table Maintenance — OPTIMIZE and VACUUM

Incremental writes create many small Parquet files over time. OPTIMIZE compacts them into larger files — Power BI Direct Lake performance depends on this. VACUUM removes old Delta file versions outside the retention window. As of March 2026, the Lakehouse Maintenance Activity in Data Factory pipelines (Preview) runs VACUUM and OPTIMIZE as a native pipeline step without a separate notebook.

-- Compact small files and co-locate data for common filter patterns
OPTIMIZE gold.sales_summary
ZORDER BY (region, category);

-- Remove Delta versions older than 7 days
-- Run during maintenance windows, not while active queries run
VACUUM gold.sales_summary RETAIN 168 HOURS;

-- Check current table health
DESCRIBE DETAIL gold.sales_summary;

Partition Strategy

Partition large fact tables by date or high-cardinality business keys so Spark and SQL endpoint queries only scan relevant partitions. Target 256 MB–1 GB per partition file after OPTIMIZE runs. Too many small partitions (under 128 MB) hurt metadata reads; too few large partitions (over 4 GB) hurt parallelism. For time-series tables, partition by year+month rather than day to avoid thousands of partitions accumulating over years.

Real-World Use Cases

Retail Analytics — Real-Time KPIs

POS data arrives via Data Pipelines Copy Job (CDC from SQL Server) into Bronze. Online orders stream through Eventstream into a separate Bronze table. Product catalogue comes from a Warehouse shortcut — no data movement. Materialized Lake Views normalise and join Bronze tables into a Silver customer-order fact. A Gold MLV aggregates sales by region, product, and channel. Power BI Direct Lake serves the Gold table — report latency matches the Eventstream ingestion rate, not a scheduled refresh window.

Finance — Power BI Premium to Fabric Lakehouse Migration

Legacy Power BI Premium datasets using Import mode are migrated to Fabric Lakehouse-backed Direct Lake semantic models via the Power BI Premium to Fabric migration path. Scheduled import refreshes are replaced by pipeline-triggered MLV refreshes — report data is as fresh as the last ETL run, not the last scheduled import. RLS moves from Power BI dataset-level to semantic model-level with Entra ID group mapping.

AI and RAG Pipelines on Fabric Lakehouse

Documents, embeddings, and structured reference data stored in a Fabric Lakehouse are accessible to Azure OpenAI via native OneLake integration — no custom ETL pipeline required to get enterprise data into the AI context. Fabric Data Agents (GA in Microsoft 365 Copilot, June 2026) can query Lakehouse tables directly using natural language, with the NL2SQL engine (Improved, June 2026 Preview) translating questions to T-SQL against the SQL analytics endpoint.

Manufacturing and IoT

Factory telemetry streams from IoT Hub through Eventstream into both a Lakehouse Bronze table and an Eventhouse KQL database simultaneously. The KQL database serves sub-second real-time dashboards with time-series visualisation (Preview, June 2026). The Lakehouse Bronze table feeds Silver and Gold MLVs for historical analysis. Predictive maintenance models trained in Fabric Data Science notebooks read from the Gold Lakehouse and write predictions back as a new Delta table, also queryable via Direct Lake.

Frequently Asked Questions – Fabric Lakehouse Tutorial