Fabric Lakehouse

Fabric Lakehouse

When to Use

• Designing a Lakehouse in Microsoft Fabric for analytics workloads
• Working with Delta tables, time travel, or schema enforcement
• Planning OneLake shortcuts to external data sources (ADLS Gen2, S3, GCS)
• Deciding between Lakehouse and Warehouse for a use case
• Preparing for Microsoft Fabric Data Engineer Associate (DP-700) exam

Core Jobs

1. Fabric Architecture Overview

• Lakehouse = OneLake storage + Delta table format + SQL analytics endpoint + Spark runtime
• All Fabric items (Lakehouse, Warehouse, Eventhouse, Semantic Model) share OneLake as single storage
• OneLake = Azure Data Lake Storage Gen2 compatible; one logical lake per Fabric tenant
• Lakehouse exposes two interfaces: Spark (read/write Delta) and SQL analytics endpoint (read-only T-SQL)

2. Lakehouse vs Warehouse Decision

Decision rule: Choose Lakehouse when you need Spark + open Delta format; choose Warehouse when your team is SQL-first and needs T-SQL DML.

3. Delta Lake Format and ACID Transactions

• All Lakehouse managed tables are Delta format (Parquet files + _delta_log/)
• ACID transactions: atomic commits prevent partial writes; concurrent reads/writes safe
• Time travel: query historical snapshots using:
  • SELECT * FROM table VERSION AS OF 10 — by version number
  • SELECT * FROM table TIMESTAMP AS OF '2024-01-15' — by timestamp
• Schema enforcement: Delta rejects writes that don't match the declared schema
• Schema evolution: mergeSchema option allows adding new columns

4. OneLake Shortcuts

• Shortcuts = logical pointers to external data; no data copy into OneLake
• Supported sources: ADLS Gen2, Amazon S3, GCS, OneLake (cross-workspace)
• Data stays at the source; access latency depends on source location and network
• Shortcuts appear as folders in Lakehouse; queryable via SQL analytics endpoint and Spark
• Use shortcuts when data must remain in source system for compliance or ownership reasons

5. Spark Notebooks in Lakehouse

• Default language: PySpark (also supports Scala, R, Spark SQL)
• Read Delta table: spark.read.format("delta").load("abfss://...")
• Write Delta table: df.write.format("delta").mode("overwrite").save("Tables/mytable")
• Shortcut: spark.read.load("Files/shortcut-folder/") — treats shortcuts as local paths
• ML feature prep: use Spark for large-scale feature engineering before model training

6. Medallion Architecture in Fabric

• Each layer can be a separate Lakehouse or schema within one Lakehouse
• Silver and Gold tables exposed via SQL analytics endpoint to Power BI (Direct Lake mode)

Key Concepts

• OneLake — single logical Azure Data Lake per Fabric tenant; all items read/write here
• Delta Lake — open-source storage layer; ACID, time travel, schema enforcement on Parquet
• SQL analytics endpoint — auto-generated, read-only SQL interface over Lakehouse Delta tables
• Direct Lake mode — Power BI reads Delta tables directly from OneLake (no import, no DirectQuery overhead)
• Shortcut — virtual link to external storage; zero data copy; Fabric-native access
• Managed table — Delta table in Tables/ section; registered in Lakehouse metastore
• Unmanaged file — raw file in Files/ section; not a Delta table; no SQL access

Checklist

• Lakehouse chosen over Warehouse when Spark or open Delta format is required?
• Medallion architecture (Bronze/Silver/Gold) layers defined for data quality progression?
• Delta tables in Tables/ section (not Files/) for SQL analytics endpoint access?
• OneLake shortcuts used for external data instead of copying data in?
• Time travel retention configured (default 7 days; extend for audit requirements)?
• SQL analytics endpoint connected to Power BI using Direct Lake mode?
• Schema enforcement enabled; mergeSchema only when intentional evolution needed?

Output Format

• 🔴 Critical — writing data to Files/ section expecting SQL access (Files are not Delta tables)
• 🔴 Critical — using Lakehouse SQL analytics endpoint for write operations (it is read-only)
• 🟡 Warning — copying data into OneLake when a shortcut could avoid duplication
• 🟡 Warning — no medallion layering; raw data directly served to Power BI
• 🟢 Suggestion — enable Direct Lake mode on Gold layer semantic model for best BI performance

Exam Tips

• Lakehouse SQL analytics endpoint = auto-generated, read-only — cannot INSERT/UPDATE via SQL endpoint; writes must go through Spark, Dataflow Gen2, or Pipelines
• Shortcuts = no data copy — data stays at source; latency depends on source; no egress costs within same region
• Delta time travel — VERSION AS OF for version number, TIMESTAMP AS OF for date; default 7-day retention
• Lakehouse tables vs Files — Tables/ = Delta format, SQL-queryable; Files/ = raw files, Spark-only
• OneLake = one copy shared across Fabric — Lakehouse, Warehouse, and Semantic Model all reference the same OneLake storage; no duplication
• Medallion architecture in Fabric — Bronze → Silver → Gold; each layer a separate Lakehouse or schema; Gold exposed to Power BI via Direct Lake