An AWS Glue ETL job builds the daily training dataset for a demand-forecasting model. Twice last quarter an upstream system shipped a bad extract, once with unit_price empty for half the rows and once with duplicate order_id values. The job loaded both extracts, and the retrained model degraded before anyone noticed. The team wants the pipeline itself to check each day's data against explicit rules and keep failing records out of the training dataset, without running a separate service or standing up infrastructure.