01 - Fabric Architecture and Workspaces¶
Microsoft Fabric overview¶
Microsoft Fabric is a SaaS analytics platform combining Power BI, Synapse Analytics, Data Factory, Data Science, and real-time streams under one roof. It is sold as capacity (compute) with OneLake (storage) as the unified substrate.
Three big mental shifts from classic Azure:
- Storage is shared. OneLake is a single tenant-wide data lake. No more per-service storage accounts.
- Compute is shared. Capacity Units (CU) cover all workloads: Spark, Warehouse, Pipelines, Power BI rendering, etc.
- Items live in workspaces. A workspace is the container for Lakehouses, Warehouses, Pipelines, Notebooks, Reports, and more.
Workspaces¶
A workspace is the unit of Fabric:
- All items (Lakehouses, Warehouses, Pipelines, Notebooks, Reports, KQL DBs) live in a workspace
- Workspace is assigned to a capacity (which provides compute)
- Workspace has roles (Admin / Member / Contributor / Viewer) for access control
- Workspaces can be Git-connected to back items in a repo
Special workspaces¶
- My Workspace - personal, no capacity assigned by default
- Standard workspace - the normal multi-user workspace (must be assigned to Fabric capacity for Fabric workloads)
Lifecycle pattern¶
Typical Dev / Test / Prod model:
Workspace: MyApp (Dev) ──┐
│ Deployment Pipeline
Workspace: MyApp (Test) ──┤
│
Workspace: MyApp (Prod) ──┘
Each stage is its own workspace; Deployment Pipelines promote items + apply per-env rules (e.g., point to a different Lakehouse per environment).
OneLake¶
OneLake is the Fabric storage. Every Lakehouse, Warehouse, KQL DB stores data here as Delta Parquet by default.
Properties¶
- One per tenant (you don't choose a region per item)
- Backed by Azure Data Lake Storage Gen2 (ADLS Gen2)
- Free / no charge for storage in trial; metered in capacity units after
- Hierarchical namespace
- Open format (Delta) - any Delta-compatible client can read
Shortcuts¶
A shortcut is a Fabric pointer to data living elsewhere. No copy. Queries fan out.
| Shortcut target | Use |
|---|---|
| Another OneLake item (Lakehouse, Warehouse) in the same or different workspace | Federated views without duplication |
| ADLS Gen2 (external) | Query existing Azure storage |
| Amazon S3 | Query AWS data |
| Google Cloud Storage | Query GCS |
| Dataverse | Query Power Platform business data |
Shortcuts honor source security (the requester needs access at the source).
Mirrored databases¶
A special form of "shortcut" that does continuous replication from operational databases:
- Azure Cosmos DB
- Snowflake
- Azure SQL DB
Mirroring lands as Delta tables in OneLake; queryable immediately, no pipeline required.
Capacity¶
Fabric is sold as capacity SKUs measured in Capacity Units (CU):
| SKU | CU | Approx. monthly (USD) |
|---|---|---|
| F2 | 2 | $263 |
| F4 | 4 | $526 |
| F8 | 8 | $1,053 |
| F16 | 16 | $2,105 |
| F32 | 32 | $4,210 |
| F64 | 64 | $8,420 |
| F128 | 128 | $16,841 |
| F256+ | ... | ... |
Pricing is approximate. F64 corresponds to a Power BI Premium P1 in the older world.
CU consumption¶
Different workloads consume CUs at different rates. A Spark notebook execution might consume X CUs/second; a Direct Lake query might consume Y CUs/second; a pipeline activity Z. The Capacity Metrics App (free Power BI app) shows per-item, per-user, per-time CU consumption.
Smoothing and bursting¶
Fabric capacity allows:
- Bursting - workloads can temporarily exceed capacity ceiling
- Smoothing - 24-hour smoothing window evens out short bursts
- Throttling - if sustained over-consumption breaches the smoothing budget, jobs get delayed/throttled
Right-size capacity by reviewing CU consumption over peak windows. Trial = F64 free for 60 days.
Free Trial¶
Sign up for a 60-day Fabric trial via Microsoft Learn. Includes a workspace assigned to free F64-equivalent trial capacity.
Items in a workspace¶
| Item | Purpose |
|---|---|
| Lakehouse | OneLake storage + Spark + read-only SQL endpoint; supports Files (semi-structured) and Tables (Delta) |
| Warehouse | OneLake-backed warehouse with full T-SQL DDL/DML |
| Notebook | Interactive Spark (Python, SQL, R, Scala) |
| Spark Job Definition | Scheduled / parameterized Spark job (no notebook UI) |
| Pipeline | Orchestration (Data Factory) |
| Dataflow Gen2 | Power Query-based low-code transformation |
| Eventstream | Streaming ingestion |
| KQL Database / Eventhouse | Real-time analytics (Kusto) |
| Semantic Model (formerly Dataset) | Power BI tabular model |
| Report / Dashboard | Power BI visuals |
| Data Pipelines | Sequence of activities orchestrating multiple items |
Items have lineage relationships - one item's output feeds another. The lineage is visualized in the workspace.
Identity and access¶
- Microsoft Entra ID (formerly Azure AD) for user identity
- Service Principals for unattended jobs
- Managed Identities when calling external resources
- OneLake security predicates for row/column-level data filtering (preview)
Git integration¶
A workspace can be linked to:
- Azure DevOps repos
- GitHub repos
Items synced as JSON / source representations. Pull requests work as expected. Promote between workspaces via Deployment Pipelines or via Git (commit to dev branch → PR to main → trigger CI).
Architecture mental model¶
Tenant
├─ OneLake (single lake, hierarchical, Delta default)
│
├─ Capacity F64
│ ├─ Workspace: data-prod
│ │ ├─ Lakehouse: bronze (raw)
│ │ ├─ Lakehouse: silver (curated)
│ │ ├─ Warehouse: gold (BI-ready)
│ │ ├─ Pipeline: nightly-elt
│ │ └─ Semantic Model: sales
│ │
│ └─ Workspace: data-dev (mirrors prod, also on F64)
│
└─ Power BI (across capacities)
└─ Reports consume gold semantic model via Direct Lake
Key exam triggers¶
- "Reference S3 data in Fabric without copying" → Shortcut to S3
- "Continuous CDC from Cosmos DB" → Mirroring (zero-ETL)
- "Promote items from Dev to Prod" → Deployment Pipelines
- "Version control for Fabric items" → Git integration
- "Single tenant-wide lake" → OneLake
- "Power BI reads Lakehouse without import or DirectQuery" → Direct Lake mode