Skip to content
Tutorials Center
My learning

Topics

Lesson 47 of 56 ยท Service, Sharing and Security

Governance, Deployment Pipelines and Best Practices

Run Power BI at scale: workspace roles, certification, sensitivity labels, deployment pipelines, naming, source control and performance guidelines.

Power BI makes it easy for anyone to build a report, which is its strength and its risk. Without agreement on who owns what, which numbers are official and how changes reach production, an organisation ends up with hundreds of near-identical reports that disagree. Governance is the set of habits and controls that prevents this, without stopping people from working.

This lesson covers workspace roles, endorsement (promotion and certification), sensitivity labels, deployment pipelines, naming conventions, source control and a performance checklist. These are features of the hosted service, Microsoft Fabric and your organisation's own policies, so everything here is described in words and any configuration is read-only. The runnable examples are small Python stand-ins that model a permission matrix and a promotion process, to make the rules concrete. Exact feature names and licence requirements change often, so check the current documentation for your tenant.

Workspace roles

Every workspace has four roles. The role decides what a person can do with all content in that workspace.

Role Typical rights
Admin Everything, including managing access and deleting the workspace
Member Edit content, publish and update apps, add members with lower roles
Contributor Create and edit content, publish reports; cannot manage access or publish the app
Viewer View and interact with content; no editing

Useful principles:

  • Give people the lowest role that lets them do their job. Most readers should not be in the workspace at all: they use an app (see Publishing to the Power BI Service).
  • Use groups (for example Microsoft Entra security groups) rather than individuals, so people joining or leaving are handled in one place.
  • Keep at least two Admins, so a departure does not orphan a workspace.
  • Remember that Admins, Members and Contributors bypass row-level security (Row-Level Security).
  • Per-item permissions, such as Build (create new content on a model) and Reshare, add finer control on top of workspace roles.

The governance toolbox

Roles Who may edit or view.
Endorsement Mark what is trusted.
Labels Classify and protect data.
Pipelines Promote changes safely.
1 / 2

Endorsement: promoted and certified

When many semantic models exist, how do people find the right one? Endorsement marks content so it stands out in search and in the Get data dialog.

  • Promoted: the owner (anyone with write access) says "this is ready to be used". It is a signal of intent, with no review.
  • Certified: an authorised reviewer, chosen by the tenant admin, verifies the item against the organisation's criteria and certifies it. Certified content is shown with a badge and appears first.
  • Master data (for data items): marks the authoritative source for a type of data such as customers.

The value of certification depends on a clear, documented checklist: owner and contact named, refresh in place and monitored, row-level security tested, naming conventions followed, documentation present, sources approved. Certify few items and re-review them, or the badge turns into decoration.

Sensitivity labels

Sensitivity labels (from Microsoft Purview Information Protection) classify content, for example Public, Internal, Confidential, Highly Confidential. A label can be applied to a semantic model, report, dashboard or dataflow and travels with the data: if you export a labelled report to Excel, PowerPoint or PDF, the label and its protection settings (for example encryption and who can open the file) can follow it.

Admins can require labels on new content, and set a default label. Labels can also be inherited: a report takes the label of its semantic model, and an exported file takes the label of the report. Labels help with compliance and with questions such as "who exported the salary data?" because activity logs show the label.

Labels do not replace RLS or workspace permissions. A label classifies and protects the file; RLS decides which rows each viewer sees.

Tenant settings and the admin portal

Some governance is not in your workspace at all but in the admin portal, where tenant admins can allow or block features: Publish to web, export to Excel or CSV, who may create workspaces or use Copilot, whether external users may be invited, the allowed data sources and more. If something described in this course is greyed out or missing, the tenant setting is often the reason. Activity logs and the usage metrics report help with audits: who viewed what, which reports nobody opens, which refreshes fail.

Administrator roles

Power BI has several kinds of "admin", and mixing them up is a common source of confusion.

Role Scope What it can do
Fabric administrator (called Power BI administrator in older documentation) The whole tenant Opens the admin portal, changes tenant settings, manages capacities, views and manages every workspace, reads the audit and activity logs, uses the admin APIs.
Global administrator Microsoft 365 / Microsoft Entra ID Includes Fabric admin rights and much more. Keep this group very small.
Capacity administrator One Premium or Fabric capacity Assigns workspaces to the capacity and manages its settings and load.
Workspace admin One workspace Adds and removes members, publishes the app, deletes the workspace (see the workspace roles above).
Report or model owner One item Controls that item's permissions, refresh and endorsement.

Give the Fabric administrator role to a few named people (at least two, so that someone is always available), never to everyone who "needs to check a setting". It is a powerful role: it can see workspace metadata across the whole organization.

Smaller admin roles split the work further. A billing administrator buys and renews licenses and capacities and sees invoices, a license administrator assigns Pro, Premium per user or Fabric licenses to people, and a user administrator manages accounts and groups. Administrators also monitor the tenant: the admin monitoring workspace and the activity log show who exported, shared or deleted what, which scheduled refreshes failed, and how capacity is used, and the admin APIs let them automate those checks and the refresh of important data. These duties belong to the IT side of self-service BI, so a report author usually only needs to know whom to ask.

Assigning an administrator role

Roles come from Microsoft Entra ID (formerly Azure Active Directory), so they are assigned there:

  1. In the Microsoft 365 admin center, open Users > Active users, select the person, choose Manage roles, and add Fabric administrator (use Show all by category if you do not see it). The Microsoft Entra admin center offers the same under Roles & admins.
  2. The person who assigns must hold a role that is allowed to do so, such as Global Administrator or Privileged Role Administrator.
  3. The new admin may need to sign out and in again before the Admin portal appears under the settings (gear) menu.

The same can be done with PowerShell through Microsoft Graph, which is useful for audits and for scripts. This block is a read-only sketch: check Microsoft's current documentation for exact module and role names before running anything like it in your tenant.

POWERSHELL
Connect-MgGraph -Scopes "RoleManagement.ReadWrite.Directory"

$role = Get-MgRoleManagementDirectoryRoleDefinition |
    Where-Object { $_.DisplayName -in "Fabric Administrator", "Power BI Administrator" }

New-MgRoleManagementDirectoryRoleAssignment `
    -PrincipalId "<user object id>" `
    -RoleDefinitionId $role.Id `
    -DirectoryScopeId "/"

The admin APIs

Administrators can read tenant-wide information with the Power BI REST API (the admin operations) and with the Power BI PowerShell module, to produce an inventory instead of clicking through the portal:

TEXT
GET https://api.powerbi.com/v1.0/myorg/admin/groups?$top=100          workspaces (add $expand=reports,datasets)
GET https://api.powerbi.com/v1.0/myorg/admin/datasets                 semantic models in the tenant
GET https://api.powerbi.com/v1.0/myorg/admin/reports                  reports in the tenant
GET https://api.powerbi.com/v1.0/myorg/admin/capacities               capacities
GET https://api.powerbi.com/v1.0/myorg/admin/activityevents?startDateTime='2026-03-01T00:00:00.000Z'&endDateTime='2026-03-01T23:59:59.999Z'
POWERSHELL
Connect-PowerBIServiceAccount
Get-PowerBIWorkspace -Scope Organization -All
Get-PowerBIActivityEvent -StartDateTime '2026-03-01T00:00:00.000' -EndDateTime '2026-03-01T23:59:59.999'

The activity events call covers one UTC day at a time and returns a continuation token for the next page. Only a Fabric administrator can use these admin calls, or a service principal that an administrator has explicitly allowed in the tenant settings (preferable for scheduled jobs, so no person's password is stored in a script). For a deeper inventory, including the lineage between sources, models and reports, the scanner APIs (workspace info) return metadata for many workspaces at once.

Try it: a workspace inventory from admin API output

The admin API returns JSON. This stand-in parses a small example in the shape of the workspaces call and answers the questions an administrator asks every quarter: which workspaces have no admin, and which are not on a capacity.

Python Stand-in: finding orphaned workspaces in admin API data

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

Two of the workspaces have no admin at all, and one has a single admin, which is a risk when that person leaves. The admin portal's Workspaces page shows the same states, and lets an administrator add themselves or someone else as admin to recover an orphaned workspace.

Creating workspaces and publishing apps

Anyone allowed by the tenant settings can create a workspace. In the service, choose Workspaces > New workspace and fill in a name (clear and unique: Sales - Production), a description, an optional image, the license mode (Pro, PPU or a capacity) and contacts. The creator becomes its first Admin. Then:

  1. Add people or groups with the right roles (prefer Microsoft Entra security groups to individuals).
  2. Publish reports and semantic models from Desktop, or create them in the browser.
  3. When the content is ready for readers, choose Create app, set its name and navigation, pick the content and the audience, and Publish app. After later changes, choose Update app. Details are in Power BI Service.

Tenant admins can restrict who may create workspaces, so that new workspaces follow your naming and capacity rules. If you cannot find a feature that this track describes, a tenant setting is often the cause.

Deployment pipelines

A deployment pipeline is a service feature that connects up to three workspaces as stages: Development, Test and Production. You build in Development, then deploy content forward to the next stage with a click (or by API). The pipeline copies reports, semantic models, dataflows and other items, and it compares stages so you can see what differs.

flowchart LR
    A["Development<br/>workspace"] -->|Deploy| B["Test<br/>workspace"]
    B -->|Deploy| C["Production<br/>workspace"]
    C --> D["App for readers"]
    A -.- E["Deployment rules<br/>swap data source and parameters per stage"]

Changes flow one way through the stages; rules adjust connections for each stage.

Two features make pipelines practical:

  • Deployment rules change settings per stage. For example, the Development model reads a small sample database and the Production model reads the real warehouse; a rule swaps the data source or a parameter (see Data Source Settings, Privacy Levels and Parameters) when content is deployed.
  • Comparison and history: each stage shows which items are new, changed or the same, and deployments are recorded.

Pipelines require a licence level that supports them (Premium per user, a capacity or Fabric, as of this writing) and workspaces assigned to a capacity. Pipelines move content, not data: each stage has its own refresh, and permissions are set per workspace, so Test can have a small group of testers and Production the app audience.

A simple release routine:

  1. Develop in the Development workspace (or publish from Desktop to it).
  2. Deploy to Test; check the numbers with a business reviewer, check RLS and refresh.
  3. Deploy to Production; update the app.
  4. If something is wrong in Production, fix it in Development and promote again; do not edit Production by hand.

Source control

A .pbix file is one binary blob, which is hard to review or merge. Newer project formats make Power BI work friendlier to version control:

  • Power BI Project (.pbip) saves the report and model as folders of text files (JSON and TMDL) that diff well in Git.
  • Git integration in the service syncs a workspace with a Git repository (Azure DevOps or GitHub, depending on the current feature set), so you can branch, commit and review changes to the model and the reports.
  • External tools such as Tabular Editor can edit the model files, and scripts can then check them.

Even without these, practise the basics: keep file names with versions or dates out of the live workspace, store the master .pbix in a shared, backed-up location, and write a short change log inside the report (for example on a hidden page) that lists what changed and why. Never keep the only copy of a production model on one person's laptop.

Naming and documentation

Consistent names remove confusion:

  • Workspaces: a team and purpose, such as Sales - Analytics (Prod), and ... (Dev).
  • Semantic models: the business subject, such as Sales Model, not Copy of Sales v3 final.
  • Tables: dimensions and facts with a clear style (Customer, Product, Sales), and no spaces or prefixes in the physical source names where possible.
  • Measures: readable, with the unit, such as Total Sales, Sales YoY %; keep them in a measure table or display folders.
  • Reports: Sales Overview, with the audience if needed.

Write down, in a wiki or in the model's description fields, the grain of each table, the definition of each key measure (what counts as a "customer", is tax included?), and the owner. A number without a definition is an argument waiting to happen.

Performance guidelines

Many "slow report" problems have the same few causes.

  • Model: use a star schema; remove unused columns and tables; avoid high-cardinality text columns, and turn off Auto date/time when you have your own date table (see Date Tables and Time Intelligence Setup).
  • Power Query: filter early, keep query folding working and use incremental refresh for big tables (see Query Folding, Incremental Refresh and Query Performance).
  • DAX: use variables, avoid iterating over huge tables when an aggregate does the job, and test measures in the DAX query view (see Debugging and Optimizing DAX).
  • Visuals: fewer visuals per page (a dozen or fewer is a good guide), avoid tables with thousands of rows, limit slicers with huge lists and use the Performance Analyzer to find the slow visual.
  • Capacity: do not refresh everything at 06:00; spread schedules out, and watch capacity metrics if you use Premium or Fabric.

Stand-in: a permission matrix

The first stand-in models workspace roles as a table of rights, and checks whether a set of people can do a set of actions. It is not the real permission system, but writing the rules down as data is a useful habit for your own governance documents.

Python Stand-in: workspace role matrix

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

Stand-in: a certification checklist

The second stand-in scores semantic models against a certification checklist and lists the ones that pass. Real certification is a human review, but the checklist structure is the same: clear, objective criteria, applied the same way every time.

Python Stand-in: certification checklist

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

Notice how the third model fails every criterion: this is the kind of "copy of a copy" that governance should find and retire. The point of a checklist is that anyone can see why something was or was not certified.

Check your understanding

Quick check ยท 4 questions

of 4 answered You scored / 4 ยท perfect!

  1. 1. What is the best way to give a large group of colleagues read-only access to finished reports?

    An app with an audience gives readers a clean, read-only view without putting them in the workspace. Using groups keeps access easy to manage.
  2. 2. What is the difference between Promoted and Certified content?

    Promotion is a self-declared signal. Certification requires an authorised reviewer to check the item against the organisation's standards.
  3. 3. What do deployment rules in a deployment pipeline do?

    Rules let each stage use its own connection, for example a sample database in Development and the real warehouse in Production, without editing the model by hand.
  4. 4. Why is the .pbip (Power BI Project) format useful?

    A .pbix is one binary file. A .pbip is a folder of text files (JSON and TMDL), so changes can be diffed, reviewed and merged in version control.

Summary

  • Give people the lowest workspace role that works, use groups, keep two Admins and share with readers through apps.
  • Promoted and Certified endorsements help people find trusted content; certification needs a documented checklist and a real reviewer.
  • Sensitivity labels classify content and can protect exported files; they complement, not replace, RLS and permissions.
  • Deployment pipelines move content through Development, Test and Production, with rules that change data sources per stage.
  • Use .pbip and Git integration for version control; keep names, definitions and owners documented.
  • Performance comes from a star schema, early filtering, simple DAX and a modest number of visuals per page.
  • These are service and organisational features: the Python examples are stand-ins for rules and checklists.

Next, put everything together in Project: Sales Performance Dashboard.