Microsoft Fabric Data Warehouse

Overview

Fabric Data Warehouse is the relational warehouse inside Microsoft Fabric, and its defining trait is that it has no storage of its own. Every table is written as Delta Parquet in OneLake, so the warehouse is a T-SQL engine over open files rather than a proprietary store.

It is developed primarily in T-SQL and shares a large surface area with the SQL Server engine, including full multi-table ACID transactions, stored procedures, functions and materialised views, which makes it the most familiar of the cloud warehouses to a SQL Server team.

There are two warehousing items and the difference catches people out. The Warehouse is read-write and accepts DDL and DML. The SQL analytics endpoint is generated automatically over a Lakehouse and is read-only, so Spark writes the data and T-SQL only queries it.

Storage and compute are separate and scaling is close to instant, but the workload management is deliberately autonomous: Microsoft's own phrasing is that there are no knobs to turn, which is the opposite of Redshift's tuning surface.

Loading is flexible, through COPY INTO, pipelines, dataflows, CTAS and INSERT..SELECT, or by pointing Spark at the same Delta tables.

Time travel is a T-SQL hint, FOR TIMESTAMP AS OF, read-only and UTC only, with retention configurable from 1 to 120 days and 30 by default, alongside CLONE TABLE for zero-copy copies.

Buying it is the part to model carefully. There is no per-query or per-node price: you buy Fabric capacity in F SKUs and the warehouse draws on the same shared pool as Power BI, Spark and Copilot.

Openness comes through OneLake's metadata virtualisation, so Iceberg tables can be read as Delta and Delta tables read by Iceberg clients, with documented limits on commit counts, latency and private links.

Key features and capabilities

The same headings are used for every cloud data warehouses entry, so two tools can be read side by side.

Architecture and storage

• A relational warehouse on a data lake foundation, with all tables stored as Delta Parquet in OneLake
• Storage and compute are separate, and the warehouse scales near instantaneously
• Two items, a read-write Warehouse and a read-only SQL analytics endpoint over a Lakehouse
• Compute is a shared pool of capacity units, not a warehouse you size

Performance and scaling

• Autonomous workload management, with no tuning knobs exposed
• Materialised views, functions and stored procedures
• A separate accelerated capacity meter exists alongside the standard warehouse meter
• Surge protection and overage controls cap and price excess consumption

SQL and workloads

• T-SQL sharing a large surface area with the SQL Server engine
• Full multi-table ACID transactions
• Loading through COPY INTO, CTAS, INSERT..SELECT and SELECT INTO, or by Spark writing the same tables
• Time travel through a FOR TIMESTAMP AS OF hint, read-only and UTC, plus CLONE TABLE
• Time-travel retention is configurable from 1 to 120 days, 30 by default

Open formats and sharing

• Delta Lake is the native format, replicated to OneLake Files for external readers
• Iceberg interoperability by metadata virtualisation, readable in both directions
• Shortcuts reach Azure storage, OneLake, Amazon S3, Google Cloud Storage and S3-compatible stores
• Published limits include Parquet only, fewer than 5,000 source commits, and no private-link support

Security and governance

• Workspace roles and item permissions, with granular T-SQL GRANT on top
• Object, column and row-level security and dynamic data masking, on both the Warehouse and the endpoint
• Microsoft Entra ID authentication as an alternative to SQL logins, and service principals through the API
• Auditing through Microsoft Purview, and customer-managed keys in Azure Key Vault
• Private Link needs a paid capacity and is disabled on the trial

AI and machine learning

• Copilot is enabled by default on paid capacities from F2 upward
• Copilot for Data Warehouse covers natural language to SQL, code completion and quick actions
• Fabric data agents keep conversation history for 28 days unless it is cleared
• Copilot is unavailable on trial capacity and in sovereign clouds, and consumes the same capacity units

Where it runs and what it costs

• Azure public cloud only, across the Americas, Europe, the Middle East, Africa and Asia Pacific
• Some regions carry Power BI only, so check the region list before committing
• No self-hosted or bring-your-own-cloud option
• Bought as capacity units rather than per query, plus OneLake storage

Pricing

SubscriptionF2 capacity at $0.36 an hour

There is no per-query price. You buy Fabric capacity at $0.18 per capacity unit hour in East US, billed per second with a one-minute minimum, which works out at $0.36 an hour for an F2 and $11.52 for an F64. The warehouse, accelerated warehouse, Copilot and OneLake meters all draw on that same rate, and overage costs $0.54 a unit hour. A one or three-year reservation saves around 41%. OneLake storage is extra, $0.026 a GB a month hot down to $0.004 cold. A 60-day trial offers F4 or F64 scale and 1 TB of OneLake.

Vendor pricing page →

Demos and videos

• Get started with data warehouses in Microsoft Fabric Tutorial · YouTube
• Introduction to Microsoft Fabric Webinar or talk · YouTube

About Microsoft

Fabric Data Warehouse is a first-party Microsoft service, sold as part of Microsoft Fabric rather than separately, so it has no founding story or funding of its own. Fabric reached general availability on 15 November 2023. Capacity is bought as Azure F SKUs billed per second with no commitment, with the option of a yearly reservation, or through a Cloud Solution Provider; the older Power BI Premium P SKUs are being retired in favour of F. It is proprietary with no self-hosted option, and the published on-ramp is a migration assistant from Azure Synapse dedicated SQL pools and SQL Server.

microsoft.com