Quickstart

This quickstart walks through the smallest useful Radar setup: one monitored workload, one SLO, and the loop that turns a threshold crossing into a governed intervention.

Prerequisites

• A Databricks workspace with a running pipeline, experiment, or model deployment to watch.
• Node.js 22 or newer and pnpm, with the fabric-radar repository cloned.
• The Fabric Platform Host configured for your organization (Radar executes every mutation through it).

1. Declare a monitor

A monitor pairs a watched workload with its SLOs — for example a cost ceiling, a freshness lag, or a quality floor. Declaring one is itself a governed action (radar.monitor_declare): it goes through schema validation and policy like any other mutation, so the watch list carries the same audit trail as the interventions it later justifies.

2. Let detection run

Telemetry flows append-only into Lakebase metrics tables and out to a Delta projection — no Platform actions involved, by design. Fabric Experiments evaluators run over Databricks system tables and MLflow on a Temporal Schedule. When an evaluator crosses a declared threshold, Radar raises radar.anomaly_raise with the evaluator evidence attached.

3. Triage the anomaly

The anomaly lands in the inbox with its evidence: which SLO breached, which evaluator fired, and the run or model version involved. A human works it from the console, or a finite Harness triage agent works it with read-only tools over the projections — its single mutation tool is the governed radar.intervention_request action, never a direct write.

4. Intervene and verify

An intervention — pause the pipeline, cancel the run, quarantine the model — executes as a governed Platform action through Harness adapters, with blast-radius policy gates in front of every write. After the adapter confirms, an evaluator checks recovery and the anomaly resolves. The loop is closed, and every step is in the event log.

Where next