Quality Alerts

What quality alerts do

A quality alert rule watches one metric of your runs. When a run finishes, AutoData compares the metric at each pipeline stage with the rule's threshold. If the comparison holds, an alert event is recorded against that run. Rules apply to every run on your account: dashboard uploads, API jobs, connector runs, scheduled runs, listeners, triggers, inference and retraining.

Evaluating the rules never affects the run itself. A run that fires an alert still completes and delivers its outputs.

Metrics

Metric	API name	Measured
Row Loss %	row_loss_pct	Percentage of the input rows no longer present after the stage. Negative when a stage adds rows, as synthetic generation does.
Null %	null_pct	Percentage of cells that are empty after the stage.
Columns Dropped	column_drop_count	Input columns minus the columns after the stage. Negative when a stage adds columns, as encoding usually does.
Duration (seconds)	duration_seconds	How long the stage took.
Drift score	drift_score	Inference and retraining runs only: how far the new data has drifted from the training session. Below 0.10 is stable; above 0.25 is a significant shift. Reported at the stage drift.

The first four are measured after each of the pipeline stages a run executes: anomaly, dtc, mdh, cds, dsm and dsg.

Creating a rule

On the Quality Alerts tab, choose New Rule:

Field	Meaning
Rule Name	A label, shown on every alert it fires.
Metric	One of the metrics above.
Op and Threshold	The comparison that fires the rule: >, <, >=, <= or ==, against a number.
Stage	Optional. Watch one stage only. Left at Any stage, the rule is checked at every stage and can fire once per stage in the same run.
Severity	Warning or Critical, to sort alerts by how urgent they are.

Rules can be edited, switched off and on with their toggle, and deleted. A disabled rule is not evaluated.

Examples

• Too many rows lost in missing-data handling: Row Loss %, > 20, stage MDH.
• Data still has gaps before generation: Null %, > 0, stage DSM.
• Scoring data has shifted: Drift score, > 0.25, stage Drift.
• A slow run: Duration (seconds), > 600, any stage.

Reading an alert

The Recent alert events list on the same tab shows when each alert fired, its severity, the stage, and a message in this form:

Quality alert: High row loss alert — row_loss_pct > 20.0 (actual: 34.5) at stage 'mdh'

The message gives the rule, the comparison, the value the run actually measured, and the stage. Each event also records the session it belongs to, so you can open that run from the Runs tab and inspect its outputs.

Notifications

Alerts are recorded on the Quality Alerts tab and through the events endpoint. A rule can store webhook_ids (shown as "Dashboard + N webhooks"), but no webhook is called when it fires, and the quality.breach event is not sent. Poll GET /api/v1/quality-alerts/events from your monitoring instead.

Fields

Key	Default	Meaning
name	required
metric	required	row_loss_pct, null_pct, column_drop_count, duration_seconds, drift_score
operator	required	>, <, >=, <=, ==
threshold	required	A number; percentages are 0 to 100.
stage	null (any)	anomaly, dtc, mdh, cds, dsm, dsg, drift
severity	warning	warning or critical
enabled	true
webhook_ids		Stored, not currently used.

API

Method	Path	Purpose
GET	/api/v1/quality-alerts	List your rules
POST	/api/v1/quality-alerts	Create a rule
PUT	/api/v1/quality-alerts/{rule_id}	Update a rule (any of its fields, including enabled)
DELETE	/api/v1/quality-alerts/{rule_id}	Delete a rule
GET	/api/v1/quality-alerts/events	The 100 most recent alert events; ?session_id= filters to one run

POST /api/v1/quality-alerts

{
  "name": "Scoring drift",
  "metric": "drift_score",
  "operator": ">",
  "threshold": 0.25,
  "stage": "drift",
  "severity": "critical"
}

The Python SDK has the same operations: client.list_quality_alerts(), client.create_quality_alert(name, metric, operator, threshold, severity="warning", stage=None), client.update_quality_alert(rule_id, ...), client.delete_quality_alert(rule_id) and client.get_alert_events(session_id=None).