Gate 1
Asset and access
Name the project, catalog table or Visual ETL job, source and target, project role, S3 and Lake Formation grants, classification, KMS, owner, and consumers.
Profile-to-enforcement evidence check
SageMaker Data Quality Readiness
Describe the table or job, access, metrics, rules, history, outputs, and response plan. Generate a checklist without treating a profile, anomaly, or quality score as root-cause or approval evidence.
Four readiness gates
Gate 2
Metrics and rules
Separate profile analyzers, learned anomaly metrics, explicit DQDL rules, severity, expected cadence, representative history, and rare legitimate conditions.
Gate 3
Outputs and response
Choose original, evaluation, row-level, failed, or passed outputs; define CloudWatch and S3 publishing, alert owner, quarantine, stop, correction, and replay.
Gate 4
Validation and lifecycle
Test known-good, null, duplicate, late, shifted, schema-changed, and evaluation-failure cases; review anomaly feedback, rule changes, cost, retention, and baseline drift.
Classify the next action
Ready to profile
Access, metrics, cadence, classification, cost, result storage, and owner are known, but learned observations are not yet used for consequential automation.
Ready to enforce
Explicit rules have fixtures, severity, output routing, stop or quarantine behavior, idempotent replay, alert ownership, and downstream acceptance evidence.
Needs investigation
An anomaly, score, or failed rule lacks source lineage, baseline context, owner, root-cause evidence, or a safe response and must not trigger irreversible action alone.
Minimum acceptance evidence
Run known-good and intentionally bad fixtures in non-production; reconcile input, passed, failed, quarantined, and written rows; verify logs, metrics, encrypted results, alert ownership, stop behavior, idempotent replay, anomaly feedback, and a consumer-approved result. Do not inject sensitive test data into broad logs.
Continue with the source guides
Profiles and anomaliesReview history, modes, feedback, score limits, and investigation ownership. Catalog table profilesSet up at-rest rules, schedules, permissions, result review, and history. Visual ETL checksDesign in-transit DQDL, outputs, publishing, stop behavior, and replay. Profiles vs rulesChoose exploration, learned change detection, explicit assertions, and operational actions.
Official facts checked August 19, 2026 against SageMaker Unified Studio and AWS Glue Data Quality documentation. Recheck permissions, supported sources, rules, modes, outputs, storage, encryption, quotas, and costs.
Frequently Asked Questions
It creates a review checklist from the asset, access, profile, anomaly, DQDL, output, response, and lifecycle facts you provide. It does not inspect AWS, run a job, read data, or change a pipeline.
Use catalog-table evaluation for stored data on demand or by schedule. Use Visual ETL Evaluate Data Quality when checks must run as data moves through a processing job.
AWS says at least three data points. Production readiness usually needs more representative history across normal, seasonal, deployment, backfill, and recovery conditions.
No. AWS says anomaly observations do not affect the score. Use explicit rules for known pass/fail requirements and investigate anomalies separately.
No. Choose block, quarantine, warn, or continue from business severity and recovery design. Test partial outputs, retries, duplicate prevention, correction, and replay before production.