SageMaker Unified Studio Data Profiling Guide
On this page
Quick answer
Amazon SageMaker Unified Studio now exposes data profiles and learned anomaly observations for stored catalog tables and data moving through Visual ETL jobs, powered by AWS Glue Data Quality.
Use profiles to learn how data behaves. Use anomaly detection to surface a metric outside its learned range. Use explicit DQDL rules to assert known requirements. None of these alone explains root cause or approves downstream use.
The operating model
| Layer | Primary output | Best use |
|---|---|---|
| Analyzer/profile | Statistics and observations | Explore shape, completeness, distribution, and change |
| Anomaly detection | Actual metric outside predicted bounds | Surface unexpected change for review |
| DQDL rule | Pass/fail against an explicit condition | Enforce known contracts and thresholds |
| Pipeline action | Continue, route, publish, or stop | Apply a reviewed operational response |
Analyzers can collect metrics without asserting a condition. Rules also collect metrics but add a Boolean expectation. This allows teams to profile first and promote stable findings into explicit controls later.
What anomaly detection learns
AWS Glue Data Quality stores profile statistics across runs. AWS says anomaly detection needs at least three data points and supports:
- Linear mode for regular cadence, trend, and seasonality;
- Fixed mode for irregular runs or flatter exploratory data.
The model predicts an expected range for a statistic. An actual value outside that range creates an anomaly observation.
That does not mean the data is wrong. A product launch, month-end load, backfill, schema migration, or legitimate seasonal event can create an unusual value. Conversely, a stable but consistently incorrect feed may look normal.
Feedback is part of the control
AWS notes that a detected anomaly becomes input to later runs unless it is explicitly excluded. Reviewers should acknowledge or reject anomalies so the baseline does not silently normalize a bad event or repeatedly alert on a valid shift.
For each observation, retain:
- table or job, metric, and run ID;
- actual value and predicted range;
- source-data window and cadence;
- deployment, backfill, or business-event context;
- reviewer decision and reason;
- excluded or accepted status;
- follow-up DQDL rule or pipeline action.
Quality score and root-cause limits
AWS says anomaly observations do not change the data quality score. A green score can coexist with an anomaly, and a failed rule can occur without a learned anomaly.
An anomaly also does not identify the cause. Investigate source freshness, upstream deployments, partition gaps, late arrivals, schema changes, duplicate loads, permissions, and business events. Preserve lineage and run logs so the observation can be traced.
At rest versus in transit
- Catalog tables: run profiles and quality evaluations on demand or by schedule against stored data.
- Visual ETL: evaluate the data as it moves through a processing job before a target write.
Use the catalog table profile guide for stored assets and the Visual ETL guide for pipeline checks.
Production checklist
- Select business-relevant metrics rather than profiling everything without an owner.
- Establish a representative schedule and at least three data points.
- Choose Linear or Fixed mode from cadence and expected behavior.
- Define reviewer and response SLAs.
- Keep known invariants in DQDL rules.
- Route metrics and detailed results to approved destinations.
- Test late, empty, duplicate, shifted, and schema-changed data.
- Review anomaly feedback and baseline drift.
Frequently asked questions
What does SageMaker Unified Studio data profiling measure?
AWS Glue Data Quality analyzers collect statistics such as row count, completeness, uniqueness, distributions, means, and standard deviations so teams can inspect change over time.
How much history does anomaly detection need?
AWS says anomaly detection requires at least three data points. The quality of the baseline still depends on representative runs, cadence, selected metrics, and feedback.
Does a detected anomaly fail the data quality score?
No. AWS says anomaly observations do not affect the data quality score. Use explicit DQDL rules when a known condition must pass or fail.
Official sources
- AWS: SageMaker Unified Studio data profiling launch
- AWS Glue: Data quality anomaly detection
- AWS Glue: DQDL reference
Source check: August 19, 2026. Recheck supported metrics, modes, history, encryption, feedback behavior, profile storage, and Unified Studio access before deployment.