Risk check before a run

Before a workload starts, Fabric Desktop by KratosBI checks what could stop the run or make it costly, and shows the evidence. The check only reads. Nothing is started, published, paused or changed to work it out.

It runs for actions that start work on a notebook, a metadata entity or a Fabric item: Run, Run in Fabric, Run now and running a schedule now. Cancel isn't checked. It's a Preview feature: its panel carries the Preview tag, and it still shows when Settings > Preview features is off, because hiding a warning would hide the reason a run may fail.

Where you see it#

Workloads. The run dialog shows the result under what the run will do. A pass folds its checks away under Show checks. Warnings and blocks are shown with their evidence, and the other checks fold away below them.
Notebook > Run in Fabric and Data mesh > Run. These have no dialog of their own, so the check stays quiet when it passes or can't be checked, and the run starts as before. With warnings or a block, a dialog shows them first. Run all in Data mesh checks each entity and shows them in one dialog. If you go ahead with some entities blocked, they still start and are refused or fail when they run.
Estate X-ray > Optimize. Its confirm always shows the check, for the policy, your sign-in and the workspace's capacity (a table has no run history, so nothing about past runs or their CU). Going ahead past warnings there isn't recorded, because a table isn't a workload.
Check again reads the capacity again instead of using the last reading.
Blocked has no way to run anyway. In a notebook whose cells don't need Fabric, Run on this PC runs it here instead. The run still reads and writes OneLake, which uses some capacity.
Warning lets you go ahead. Going ahead is recorded (see below) (except Estate X-ray > Optimize, above). When the settings ask for it, you tick I've looked at these and want to run anyway first.

A notebook not yet linked to its Fabric item isn't checked, and its first Run in Fabric starts as before.

The engine itself reports only: starting a run (workload_act) doesn't check it, and results say "enforced": false. A run started another way, such as from an AI tool or the local API, isn't stopped by a blocked result. The dialogs above are where a block stops a run.

Results#

Result	Means
Pass	Nothing found that should stop or slow the run.
Warning	Something to look at first, such as a recent failure, a run already going, or a CU threshold you set. You can go ahead.
Blocked	An actual restriction: the run would be refused or fail. Examples are your organization's policy, not being signed in, a paused capacity, a problem in the definition, or the memory guard on this PC. A cost estimate never blocks a run.
Unknown	The capacity's state couldn't be read and nothing else was found. The run may still work.

Each finding has a title, what it means, the evidence, where the evidence came from and what to do. Findings are listed worst first.

What is checked#

Check	Source	Outcome
Organization policy	`allow_fabric_runs` (policies)	Blocked when IT turned Fabric runs off. Runs on this PC aren't affected.
Sign-in	This PC's sign-in, without a prompt	Blocked when the action uses Fabric and you're signed out.
Capacity	The capacity reading: GET `/workspaces/{id}` and GET `/capacities` with your sign-in	Paused on a reading from the last hour: blocked. Paused on an older reading: warning, with its age. Running: pass, but how busy it is isn't known, because Fabric has no utilization API for developers. Not visible to you, or Capacity.Read.All not granted: unknown, with what to ask for. No workspace, no capacity, or no access to the workspace: blocked.
Definition	The project's definitions and deployment records	Problems in the definition: blocked. An environment running an older version: warning.
Notebook compatibility	The same check as Runs on	Cells that won't work where it runs: warning. For a Fabric run, whether this PC could run it instead. A notebook that needs Fabric is never suggested for this PC.
Input size	The source's Delta log (metadata entities)	For reference in Fabric. On this PC, input over the memory guard: blocked.
History	Runs of this workload recorded by this app	The last run failed, or too many of the last five failed: warning. A run still going: warning. Usual run time, for reference.
CU use	Measured usage matched to this workload's runs, otherwise the app's estimate	For reference, labeled measured or estimated. Above the threshold you set: warning. With neither, it stays unknown, never 0.
Schedules	Schedules on this PC and in Fabric	A schedule on this PC due soon: warning. Fabric schedules' next run times aren't read.

The check makes at most two reads in Fabric: the workspace and the capacity list. It uses an environment's capacity reading from the last hour when there is one. Each result lists the reads it made.

Settings#

Change them in the Cost page's Prices dialog, under Before a run, or with set_risk_settings. The settings are kept with this PC's settings.

Setting	Default	Effect
`warn_run_cu_hours`	off	Warn when the last run of this workload used more CU hours than this. Measured when an import matched the run, otherwise estimated, and labeled so.
`warn_failed_runs`	2	Warn when this many of the last five runs failed. 0 turns it off.
`schedule_window_minutes`	60	Warn when a schedule on this PC runs the workload within this many minutes.
`require_acknowledgement`	off	The run dialogs ask you to tick "I've looked at these" before running with warnings.

Going ahead with warnings#

risk_acknowledge records that you saw an assessment's warnings and chose to run anyway. It saves who, when, the warnings and an optional note in risk_acknowledgements.jsonl in the app's data folder, and the call is in the audit trail.

The assessment is checked again first:

A blocked result can't be acknowledged.
If the findings changed since you looked, the acknowledgement is refused and you need to look again.

Checking is never audited, because it only reads.

For AI tools#

The fabric_desktop_check_run_risk capability (Preview) returns the same result to tools connected through MCP or the local API. It reads Fabric with your own account.

Limits#

The capacity's state says whether work can start, not whether it will be slow: utilization isn't available to developers.
Matching CU to this workload uses the runs this app started. Runs started elsewhere count only through imported measured usage.
No CU number is predicted for the next run. The last run's figure is shown for reference, labeled measured or estimated.
Not yet validated against a live tenant. Tests use fixtures.