Fabric Workloads and Core Objects

Read the organizational hierarchy

Before creating anything in Fabric, it pays to know which box you are creating it in. Almost every administrative question — who can see this, what pays for it, where does it live — is answered by naming the right level of the hierarchy.

A tenant is one Fabric instance for an organization and is aligned with a Microsoft Entra tenant. One organization, one tenant, one unified OneLake underneath it. Above the workspace level, an organization can group workspaces into domains, such as Finance or Marketing, so administrators can delegate management and apply governance policies to a business area.

Inside the tenant, a capacity supplies a dedicated set of resources that can be used at a given time. Capacity defines the ability of a resource to perform an activity or to produce output, and Fabric offers it through Fabric SKUs and trials.

The meter that goes with it is the capacity unit (CU), the unit of compute measurement in Fabric. All workloads and operations consume capacity units from the available capacity, and that consumption is classified as interactive or background. Different items consume different amounts at a given moment.

Work inside a workspace

Teams do not build directly in the tenant. A workspace is a collaborative container for Fabric items and uses capacity for its executed work. It brings different functionality together in one environment designed for collaboration, and it provides controls over who can access the items in it.

An item is an object created in Fabric, such as a lakehouse, notebook, warehouse, pipeline, or eventhouse. Users create reports, notebooks, and semantic models inside a workspace, and the work those items execute draws on the assigned capacity.

Read as a ladder, the hierarchy is: tenant, then capacity-backed workspace, then items inside the workspace. Getting the order right also rules out the shapes that sound plausible but are inverted, such as a tenant living inside an item or a capacity living inside a notebook.

So when an administrator needs collaboration boundaries, access controls, a place to group lakehouses and notebooks, and compute drawn from assigned resources, the answer is a single object with a single setting: a workspace assigned to Fabric capacity.

A workspace is also none of the following: it is not a Spark application split into parallel jobs, it is not a stream catalog without user-created objects, and it is not Direct Lake, which is a storage mode that lets semantic models read Delta tables straight from OneLake without importing data or using DirectQuery.

Separate workloads from items

Here is the distinction that trips up almost everyone in week one. A workload is a collection of capabilities for a particular function; an item is an object created with those capabilities.

Put plainly: Data Factory is not a pipeline, and Data Engineering is not a notebook. The workload is the shelf; the item is what you take off it.

A pipeline is a Data Factory item used to orchestrate data movement and transformation. A notebook is a Data Engineering item used to author code, run and monitor Spark jobs, view results, and collaborate. Each Fabric workload includes different item types, which is why naming the workload first usually narrows the choice of item to two or three candidates.

Note what neither of them is. A workload does not store files, and an item does not measure compute — that is what capacity units do. A workload is not a tenant, and an item is not a security principal.

Recognize the engineering objects

Data Engineering exists for large-scale processing with Apache Spark, and its first item is where the data sits. A lakehouse combines files, folders, and tables over a data lake and supports both Spark and SQL engines for big-data processing. Its Delta-formatted tables support ACID transactions, and it holds structured and unstructured files organized in folders and subfolders inside a workspace folder in OneLake.

Delta is not an exotic choice here: Delta Lake is the standard table format across all Fabric workloads, so data ingested into Fabric is stored as Delta tables by default.

The second item is the batch worker. A Spark job definition is a set of parameters that describes how a Spark application should run — input and output data, transformations, and configuration — and it lets you submit batch or streaming jobs to the Spark cluster.

Use the notebook as the interactive surface

The third engineering item is the one engineers spend their day in. A notebook is a multi-language interactive programming tool that supports authoring code and Markdown, running and monitoring Spark jobs, viewing results, and collaborating with team members. It serves exploration, preparation, and machine learning experiments alike.

Its second life matters as much as its first: a notebook can be transformed into a pipeline activity for orchestration. Interactive work during development, scheduled execution in production, one artifact throughout — no rewrite into a different item when the code is ready to run on its own.

Tell the two kinds of pipeline apart

Data Factory supplies the movement and orchestration items, and one of its names is badly overloaded. A Data Factory pipeline orchestrates data movement and transformation and is different from a Fabric deployment pipeline.

The confusion is only in the word. The first is an item you build a data solution with; the second belongs to release management. Neither of them measures capacity units, defines a tenant, supplies an identity, or produces a report.

Alongside pipelines, Dataflow Gen2 is the low-code Power Query interface for ingesting data from hundreds of sources and transforming it. A connector attaches Data Factory to a type of data store, and a trigger starts pipelines on conditions such as a schedule or the arrival of data.

These objects may cooperate with Data Engineering objects in the same workspace: a Dataflow Gen2 can prepare data with a visual interface while a pipeline sequences the activities that surround it, and the pipeline can call a notebook as one of those activities.

Locate shared data in OneLake

OneLake is the single storage layer for Fabric workloads, so the last group of objects is about reaching data rather than holding it.

A shortcut is an embedded reference within OneLake that points to another file store location, enabling access to external data without copying it or building an ETL pipeline. The data stays where it is; only the reference lives in OneLake.

A semantic model is a metadata layer of tables, relationships, measures, and data connections that supplies data to Power BI reports, dashboards, and other analytics experiences.

Keeping these two apart is easy once their jobs are stated side by side, and it also separates them from objects that sound similar:

Key takeaways

• A tenant contains capacity-backed workspaces, and workspaces contain items. Capacity units measure the compute those items consume.
• Workloads are capability collections; items are the objects users create. Data Factory is not a pipeline, and Data Engineering is not a notebook.
• Core engineering items include lakehouses, notebooks, Spark job definitions, pipelines, and Dataflow Gen2. A notebook can also run as a pipeline activity.
• A Data Factory pipeline orchestrates data work; a deployment pipeline promotes Fabric content between stages.
• Shortcuts reference external data without copying it, and semantic models supply modeled data to reports.