Semantic model best practices (Preview)

KratosForge checks a Data mesh semantic model (SM_<Layer>) against Best Practice Analyzer (BPA) rules by running semantic-link-labs in Fabric. semantic-link-labs is MIT licensed. It needs Microsoft's semantic-link library, which runs only inside Fabric, so the check runs there and not on this PC.

The engine is in place. The screens (a Best practices card on Data mesh > Deploy > Semantic model, and model rows in Optimize > Opportunities) come next, once the mockup is approved.

What a check does#

It finds the model by name in the environment's workspace. The model must already exist, so create or update it first.
It runs the risk check for a run in Fabric in that workspace. A blocked result stops the check. A warning goes ahead only after the person acknowledges it, and the acknowledgement is kept with the others (risk_acknowledgements).
It publishes the notebook nb_fabric_desktop_bpa to the workspace, unless the same notebook is already there. It is a Fabric Python notebook whose default lakehouse is the environment's LH_Meta.
It runs that notebook, with inline installation turned on for the run (_inlineInstallationEnabled), since a notebook run as a job skips %pip otherwise. The notebook installs semantic-link-labs 0.17.1 into its own session, then calls run_model_bpa(..., extended=True, return_dataframe=True). extended collects the model's VertiPaq statistics, which some rules need. The model is read through its XMLA endpoint, so the capacity needs XMLA read, which is on by default.
The notebook saves the findings to LH_Meta/Files/_fabric_desktop/results/<request>.json. The app reads them back once the job finishes and keeps the last result for each model on this PC.

What it writes: the notebook item (step 3), one result file per check (step 5), and a line in the app's Fabric runs, so the check shows in Monitor. What it never changes: the model, its reports or its data.

Capacity: every check runs a notebook on the workspace's capacity, so it uses capacity. It is a Python notebook, not Spark, to keep that small.

Rules#

The rules are semantic-link-labs' own set: 64 rules in 0.17.1, ported from Microsoft's BPA rules. Findings come back with a severity (Error, Warning or Info), a category, the rule, the object, and what the rule asks for.

The generated model already follows the rules that KratosForge controls. For example, no column sums on its own, and each fact has a total measure per numeric column instead. Findings that the generator causes are fixed by the next Update model, not by editing the model in Fabric.

Engine methods#

Method	What it does	Changes
`model_bpa_status(env, layer?, name?)`	The model's last findings, a check still running, or one that failed in Fabric	Nothing in Fabric
`model_bpa_risk(env, layer?, name?, refresh?)`	The risk check for checking this model in Fabric	Nothing
`model_bpa_check(env, layer?, name?, risk_id?)`	Starts a check	Publishes the notebook and starts a Fabric job (capacity)

Permissions#

There is no new sign-in scope. The notebook runs as the person who starts it, so they need to be able to run notebooks in the workspace and read the model. See permissions.