Post

Fabric Key Concepts (first principles explanation)

Posted
By Shoaib Akthar
6 min read
Fabric Key Concepts (first principles explanation)

Instead of memorizing definitions, think about the fundamental problem each component solves.

1. Lakehouse

Where do we store huge amounts of data cheaply while keeping it easy to analyze?

Traditionally there were two choices:

SystemStrengthWeakness
Data LakeCheap, flexibleSlow analytics
Data WarehouseFast analyticsExpensive, rigid

The idea behind a Lakehouse is:

Can we combine the flexibility of a data lake with the analytical capabilities of a warehouse?

A Lakehouse stores files directly in cloud storage but organizes them in a structured way.

1
2
3
4
5

Cloud Storage
│
├── sales.parquet
├── customers.parquet
├── inventory.parquet

But with metadata, schemas, transactions and query optimization.

In Fabric:

Lakehouse = scalable file storage + tables + Spark processing

Used for:

  • raw data ingestion
  • data engineering
  • machine learning
  • big data workloads

For example:

1
2
3
4
5
6
7
8
9

ERP Data
CRM Data
API Data
↓
Fabric Lakehouse
↓
Transformations
↓
Analytics

2. Warehouse

How do we make business reporting extremely fast?

Business users need:

  • SQL
  • dashboards
  • aggregations
  • joins
  • predictable performance

A warehouse is designed specifically for that.

1
2
3
4
5
6

Fact Sales
Fact Orders

Dim Customer
Dim Product
Dim Date

Everything is optimized for:

1
2
3

SELECT SUM(Sales)
FROM FactSales
GROUP BY Region

Fast execution.

Warehouse sacrifices flexibility for speed.

In Fabric:

Warehouse = SQL-first analytical database

Good for:

  • Power BI
  • reporting
  • star schemas
  • dimensional modeling

Lakehouse vs Warehouse

LakehouseWarehouse
File basedDatabase based
Spark friendlySQL friendly
FlexibleStructured
ML workloadsBI workloads
Semi-structured dataRelational data
Cheap storageQuery optimized

Lakehouse:

Store everything.

Warehouse:

Store curated business data.

3. OneLake

Why do organizations create hundreds of copies of data?

Traditionally:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17

Sales DB
↓
Copy
↓
Data Lake
↓
Copy
↓
Warehouse
↓
Copy
↓
Power BI Dataset
↓
Copy
↓
ML Platform

Lots of duplication.

Problems:

  • storage cost
  • synchronization
  • governance
  • version issues

Microsoft asked:

Why not have one storage layer for everything?

That’s OneLake.

Think of it as:

1

OneDrive for enterprise data

Everything lives here.

1
2
3
4
5
6

OneLake
│
├── Lakehouse
├── Warehouse
├── Eventhouse
├── Semantic Models

Single source of truth.

One copy of data.

Multiple consumers.

4. Semantic Model

Why can’t business users query raw tables directly?

Raw tables:

1
2
3
4

customer_id
cust_key
sales_amt
prod_id

Business users think differently.

They think:

1
2
3
4
5
6

Revenue
Profit
Customer
Year
Margin
Growth

Semantic layer translates technical data into business concepts.

Example:

Raw:

1

SUM(order_value)-SUM(cost)

Semantic model:

Profit =
SUM(Sales[Revenue])
-
SUM(Sales[Cost])

Users simply see these in Power BI visuals, with the ability to view by different dimensions.

1
2
3

Profit
Region
Month

Semantic Model contains:

  • measures
  • relationships
  • hierarchies
  • business logic
  • security

Think:

Semantic Model = business language built on top of data.

It answers:

What does revenue mean in our company?

5. Pipeline

How do we automate movement and transformation of data?

Without pipelines:

1
2
3
4
5
6
7

Man downloads CSV
↓
Runs notebook
↓
Runs SQL
↓
Refreshes reports

Manual. Not scalable.

Pipeline is orchestration.

Think:

1
2
3
4
5
6
7
8
9
10
11
12
13

Step 1:
Copy Data
↓
Step 2:
Run Notebook
↓
Step 3:
Execute SQL
↓
Step 4:
Refresh Semantic Model
↓
Done

Pipelines answer:

What runs?

In what order?

When?

Example:

1
2
3
4
5
6
7
8
9

Daily at 2 AM
↓
Load ERP Data
↓
Clean Data
↓
Update Warehouse
↓
Refresh Power BI

Pipeline ≠ transformation.

Pipeline = coordinator.

Like a movie director.

6. Notebook

How do data scientists and engineers write flexible code close to the data?

Notebook provides the ability to write complex logic in code:

1
2
3
4
5

import pandas as pd

df = spark.read.table("sales")

df.groupBy("Region").sum()

Notebook combines:

  • code
  • outputs
  • charts
  • documentation

in one place.

Supported languages:

  • Python
  • PySpark
  • Scala
  • SQL

Notebook = interactive workspace for experimentation.

Data Scientists use notebooks because they can:

1
2
3
4
5
6
7
8
9

Write code
↓
Run immediately
↓
Inspect results
↓
Modify
↓
Run again

7. Spark

What happens when data no longer fits into one computer?

Suppose you have 5 TB dataset.

Laptop cannot process it.

Solution: Split work across many machines.

Spark does this.

Example: 100 GB file. Spark divides it.

1
2
3
4
5
6
7
8
9
10
11

Machine 1
25 GB

Machine 2
25 GB

Machine 3
25 GB

Machine 4
25 GB

All process simultaneously.

Results merged.

Spark provides:

  • distributed processing
  • fault tolerance
  • scalability
  • in-memory computation

Think:

Spark = engine for parallel computing.

Without Spark:

1

One worker

With Spark:

1

100 workers together

Fabric Lakehouse heavily uses Spark.

8. Eventhouse

What if data arrives continuously instead of once per day?

Examples:

  • IoT sensors
  • clickstreams
  • application logs
  • telemetry
  • stock ticks

Traditional systems assume:

1

Batch data

Eventhouse assumes:

1

Data never stops arriving

Example:

1
2
3
4
5
6
7
8
9
10
11

Temperature Sensor
↓
22.1
↓
22.3
↓
22.5
↓
22.4
↓
22.6

Thousands every second.

Eventhouse is optimized for:

  • streaming ingestion
  • real-time analytics
  • time-series queries
  • logs

Warehouse:

What happened last month?

Eventhouse:

What is happening right now?

9. KQL

KQL stands for: Kusto Query Language

SQL is good for tables, but is it ideal for logs and streaming data?

Consider logs:

1
2
3

2026-06-28 10:01 Login Success
2026-06-28 10:02 API Error
2026-06-28 10:03 Login Failed

Searching patterns with SQL becomes cumbersome.

KQL is designed for this.

Example:

Logs
| where Status == "Error"
| summarize count()
by bin(Timestamp,1h)

Meaning:

1
2
3
4
5

Find errors
↓
Group by hour
↓
Count them

KQL excels at:

  • telemetry
  • monitoring
  • anomaly detection
  • security logs
  • event analytics

Common with:

  • Eventhouse
  • Real-Time Intelligence

  • Azure Data Explorer

    The Entire Fabric Picture

1
2
3
4
5
6
7
8
9
10
11
12
13
14

            OneLake
               │
 ┌─────────────┼─────────────┐
 │             │             │
Lakehouse  Warehouse   Eventhouse
 │             │             │
Spark        SQL           KQL
 │             │             │
Notebook   Semantic Model  Real-time Analytics
        \       │       /
         \      │      /
            Pipeline
               │
         Power BI Reports
ComponentThink of it as
OneLakeHard drive for the entire company
LakehouseFlexible data repository
WarehouseOptimized SQL database
Semantic ModelBusiness vocabulary layer
PipelineWorkflow orchestrator
NotebookCoding workspace
SparkDistributed compute engine
EventhouseReal-time data store
KQLQuery language for events and logs

That’s it. You now know the most important concepts in Microsoft Fabric. And most importantly you got to know why we need each component with proper understanding.

Fabric
Fabric DP-600
This post is licensed under CC BY 4.0 by the author.