Skip to content
Manuel Luján edited this page Nov 6, 2026 · 1 revision

FAQ

Why not just use Microsoft Purview?

Purview is the official governance catalogue and you should keep using it. For Fabric items other than Power BI, though, Purview documents only item-level lineage: it does not support cross-workspace lineage or notebook-to-pipeline lineage for those items, and it cannot tell you which tables a notebook reads or writes. LineageIQ fills that gap at object level; it does not replace Purview. See How it compares for the details and sources.

Is it safe to connect to my tenant?

Discovery is read-only by design: the scripts and notebooks read from the Fabric and Power BI APIs and write only LineageIQ's own results. Other safeguards:

  • Reads need a Microsoft Entra ID sign-in or an admin-issued API key, and admin routes check an allowlist that is empty by default.
  • API keys are stored only as SHA-256 hashes, can expire, and are read_only unless an admin chooses full.
  • The Cloudflare quick tunnel is off by default, and Docker Compose publishes the port on 127.0.0.1 only.
  • Demo mode serves reads without sign-in and is meant only for the synthetic Contoso data.

The security policy describes the security model and how to report a problem.

How does it get lineage out of code?

It parses notebook code, pipeline definitions and semantic model TMDL to find reads and writes of tables, views, stored procedures and files. The parsers use regular expressions and structured JSON traversal, not an AI model. Object types are then confirmed against SQL endpoint catalogues where it can reach them (for example NB_Extract_SQL_Endpoint_Catalog and NB_Confirm_Source_Object_Types). Each edge carries a confidence level, confirmed, likely or unknown, and review queues list the uncertain cases.

Does it scale to a large tenant?

It is designed for tenant-wide scans: it can run inside Fabric as a pipeline of notebooks on your own capacity, and scans started from the web UI are queued (up to 20 requests) while one is running. The Power BI Scanner API caps PostWorkspaceInfo at 500 requests per hour, so on a large tenant schedule the governance scans instead of running them back to back.

Why is there no column-level lineage?

Doing it reliably means analysing SQL and PySpark in depth, and the project prefers not to promise results it cannot back up. Columns are catalogued and profiled, but lineage edges stop at object level. Column-level lineage is listed as a non-goal on the Roadmap for now.

Do I need Fabric access to try it?

No. The demo image runs on a synthetic Contoso tenant with no Fabric access and no credentials.

Can I use it at my company?

Yes. It is released under the MIT License. Note that it is an early-stage project (version 0.2.0) and, until 1.0.0, minor versions may change behaviour.

Do I need an Entra app registration?

Not to start. When entra_client_id is empty, the device-code sign-in uses the Azure CLI's public client ID. Fabric-native discovery does need a service principal; see Fabric-native discovery setup.

What permissions does discovery need?

  • Local scripts: an Entra ID account that can read the target workspaces. A scan sees only what that identity can see.
  • Fabric notebooks: a service principal with the Contributor role on LineageIQ's own workspace; for tenant-wide discovery it also needs the Fabric Administrator role.
  • Governance panel (sensitivity labels, endorsement, tags, RLS roles from the Scanner API): extra admin permissions, listed in the admin panel's "Tenant & auth" tab.

The admin panel and authentication guide has the details.

Does it read personal data?

It does not index row-level personal data. The Privacy panel is a locator: it can say a table might hold a category of personal data, never return a person's row. Name-based sensitivity matching reads only column names. The content-based check reads a bounded sample of values and stores match percentages; storing real sample values is off by default (STORE_REAL_SAMPLE_VALUES = False).

Which AI clients can use the MCP tools?

Any MCP client. The stdio servers work with clients that launch a local process, such as Claude Desktop and Claude Code. The Streamable HTTP endpoint at /mcp works with clients that connect by URL with a custom header; the remote MCP setup guide covers Claude Code, Claude Desktop, ChatGPT, Cursor, Windsurf and GitHub Copilot.

How is it tested?

At v0.2.0, CI runs 423 Python unit tests that need no live Fabric connection, byte-compiles the Python sources and syntax-checks the JavaScript. The demo image workflow also starts the image and checks the web UI, the demo data and the MCP server before each release is published. Playwright and Pester suites exist for local runs.

Is the UI available in other languages?

The UI starts in English and has a Spanish switch, with light, dark and automatic themes.

Where do I ask something that is not here?

In Discussions, Q&A.