Previous version: MLA-C01. The current version is
MLA-C02.
Choosing a data-validation tool before training
| Capability | Glue Data Quality | Glue DataBrew | SageMaker Clarify (pre-training) |
|---|
| Primary job | Rule-based quality checks (DQDL) | Visual profiling + no-code cleaning | Bias detection (CI, DPL, and more) |
| What it answers | Are values correct and complete? | What does the data look like, and clean it | Is a facet under-represented or mislabeled? |
| How rules are defined | DQDL rules, auto-recommended by profiling column data | Point-and-click profiling and transforms | Configure facet, label, and metric set |
| Built on / output | Open-source DeeQu; data-quality score (% rules passing) | 250+ built-in transforms; profile report | Model-agnostic metrics on the raw dataset |
| Best fit | Automated quality gates in ETL / Catalog | Exploratory cleanup without writing code | Fairness audit of training data before training |
Decision tree
Cheat sheet Cheat sheet
- SageMaker Clarify measures bias on the dataset before training, model-agnostically
- Class Imbalance (CI) measures sample counts and ranges from -1 to +1
- DPL measures uneven positive-outcome rates across facets
- Clarify reports eight pre-training bias metrics, with CI and DPL the headline pair
- Pre-training bias is Clarify; bias drift in production is Model Monitor
- Fix a starved minority class by rebalancing: resample or generate synthetic data
- SMOTE-style synthetic data interpolates new minority points, not duplicates
- Rebalance before you split, and never resample across the train/test boundary
- Split into train/validation/test and shuffle to prevent order leakage
- AWS Glue Data Quality uses DQDL rules and reports a data quality score
- Glue Data Quality: rule recommendations from the Catalog, failing-record detection from ETL
- AWS Glue DataBrew is visual, no-code profiling and cleaning
- Encrypt ML data at rest with S3 SSE-KMS and in transit with TLS
- Amazon Macie discovers sensitive data in S3 using managed data identifiers
- Anonymize or mask identifiers before they reach the feature set
- Data residency fixes the Region for the bucket and training resources
- Glue Data Quality dynamic rules adapt thresholds to history with last()
- Glue Data Quality static rules gate loads; Fail-without-loading stops bad data
- A DataBrew profile job generates data-quality statistics with no code
- Model Monitor baselines the training data, then reports drift in constraint_violations.json
- DataCaptureConfig samples and encrypts endpoint traffic for Model Monitor