Skip to content
TechFabric

Workload observability

Know what your Databricks estate is doing, and who may touch it

Governed mission control for running data and ML workloads on Databricks.

3 common questions, answered below ↓
Role in the family
Workload observability
Databricks surfaces
Databricks JobsLakebaseUnity CatalogTemporal
Status
Domain module, policies, state machines, and org-scoped projections are implemented, with a Temporal intervention worker and a Databricks effector.

The problem

Monitoring tells you a job failed. It rarely tells you whether that breach matters, who is allowed to intervene, or what was actually done about it last time. The intervention itself usually happens in a terminal, unlogged.

How it works

Monitors carry declared service level objectives. When reality breaches one, Radar triages the anomaly and makes every intervention a governed action with a policy check and an audit record. Telemetry ingest is deliberately append-only and stays outside the governed pipeline; only state transitions pass through it.

What it does

SLOs as declarations

What 'healthy' means is written down per workload.

Governed intervention

Restarting a job or rerouting a pipeline is an action with a policy and an audit trail, not an untracked click.

Durable remediation

Interventions run as Temporal workflows, so a long remediation survives a restart mid-flight.

Anomaly triage

Detection composes the same evaluation logic.

What it changes for you

Someone is watching your live workloads, and every intervention is on the record. Declared SLOs replace tribal knowledge about what healthy looks like, and a two in the morning fix becomes a governed action with an audit trail.

Where this shows up in an engagementMigrations to Databricks

The job failed. Nobody knows whether it mattered.

Monitoring answers whether something broke. The questions that actually cost money are the three it does not answer.

An alert with no severity

A job failed. Whether that breach matters, and to whom, is worked out by somebody who happens to know what that job feeds.

Nobody is allowed to act

The person who sees the alert cannot stop the workload, and the person who can is not looking. The gap between those two is where the bill accrues.

The bill arrives without a cause

Spend doubled last quarter. Which jobs did it, and whether any of them produced anything, is a forensic exercise rather than a dashboard.

All three are the same missing thing. Observation is not the hard part. Deciding what a breach means and who may respond to it is.

The shift

Three ways to watch a live estate

Only one of them lets somebody act on what they see.

Platform monitoring

It ran, or it did not

Job runs
Metric emitted
Alert fires
Triage by hand
Guess at impact

A dashboard somebody built

Accurate, and unread

Metrics collected
Dashboard drawn
Nobody watches
Discovered later
Explained afterwards

Governed mission control

Watched, scored, and actionable

Signal
Severity by policy
Intervention offered
Approved by someone allowed
Recorded

What gets installed

Observation, judgement and the authority to act, as one system.

Workload signals

Job state, latency, cost and data quality collected as events rather than as scraped dashboard state.

Severity from policy

Whether a breach matters is evaluated against rules somebody wrote, not inferred by whoever is on call.

Interventions with authority

Stopping a workload is an action with a policy in front of it, so the person who sees the problem can be the person who fixes it.

Cost, attributed

Spend traced to the jobs and the teams that caused it, so a growing bill has a name rather than a mystery.

Budgets that act

A budget that stops a workload rather than emailing about it after the quarter closed.

A record of every response

What fired, who acted, under which policy. Reading the estate's history is a query.

Policy

Decides severity, rather than whoever is on call

Built for estates we were on the hook for

Radar came out of running Databricks workloads where somebody was going to ask us what happened and why. It is free and open source under Apache-2.0 and it runs in your workspace.

FAQ

TechFabric Radar, answered

How is this different from the monitoring we already have?

Monitoring tells you a job failed. Radar carries declared service level objectives per workload, so it can tell you whether the breach matters, who is allowed to intervene, and what was actually done about it last time. If you are not sure the workspace is worth that yet, the Databricks Health Check at /databricks/health-check is the two-week way in.

What does a governed intervention mean in practice?

Restarting a job or rerouting a pipeline runs as a policy-checked action with an audit record, executed as a durable workflow. A fix at two in the morning stops being an untracked click in a terminal.

Does it ingest all our telemetry?

Telemetry ingest is deliberately append-only and stays outside the governed pipeline. Only state transitions pass through it, which keeps the audit trail meaningful instead of drowning in events.