What is Fabric Radar?
A governed mission control for running data and ML workloads on Databricks — monitors with SLOs, evidence-backed anomalies, and interventions that are gated, audited, and verified.
Fabric Radar is a governed mission control for running data and ML workloads — pipelines, experiment runs, model deployments, and the cost, quality, and freshness anomalies they produce — built as a Fabric family vertical on Databricks. It watches what is live, detects when reality breaches declared SLOs, and makes every intervention a governed, auditable action.
Radar is deliberately thin. Your Databricks workspace stays the system of record for telemetry and truth: system tables, MLflow, and Unity Catalog are the sources Radar reads. Radar adds the three things a workspace does not give you on its own:
- Declared monitors with SLOs — a watched workload plus its cost ceilings, freshness lags, and quality floors, declared as a governed action so the watch list itself is audited.
- Evidence-backed anomalies — Fabric Experiments evaluators over system tables and MLflow cross a threshold and raise an anomaly with the verdict attached, into an inbox humans or a finite triage agent can work.
- Governed interventions — pause a pipeline, cancel a run, or quarantine a model through Fabric Platform actions with blast-radius policy gates, durable execution, and an evaluator that confirms recovery before the anomaly resolves.
Every state transition is a Fabric Platform action invocation. Raw telemetry is the one deliberate exception — it flows append-only and never becomes a mutation path.
Start here
Quickstart
Declare your first monitor over a live workload and watch the detection loop start.
Architecture
Monitors, anomalies, and interventions — the concepts and the loop that connects them.
Deployment
How Radar's docs, console, and control plane are deployed across Cloudflare and Databricks.
What Radar owns — and what it delegates
Radar owns the monitoring domain: the monitor registry, the anomaly lifecycle, the intervention action definitions, the projections, and the console. It delegates everything else to the Fabric product that owns it.
| Responsibility | Product |
|---|---|
| Governed mutations, policy, state machine, audit, recovery | Fabric Platform |
| Databricks transport, Temporal intervention workflows, triage agents | Fabric Harness |
| Evaluators over system tables and MLflow, anomaly evidence | Fabric Experiments |
| Delivery and promotion of Radar itself | Fabric Runway |
| Monitoring domain, anomaly inbox, intervention definitions, console | Fabric Radar |
Harness builds the agents. Experiments judges them. Runway ships them. Tower flies the missions. Radar watches everything in the air. Platform keeps everyone honest.
Radar does not implement a second orchestrator, an anomaly-detection engine, a metrics TSDB, or a bespoke approval queue. Temporal stays the runtime of record; Radar signals it, never replaces it.
Tower steers agents; Radar steers workloads
Fabric Tower is mission control for human-steered agent squads; Radar is mission control for machine-run workloads. They share the Platform runtime, the event log, and the console shell — nothing else. When a Tower mission needs to act on a workload, its agent calls Radar's governed intervention actions; Tower never grows its own intervention machinery.