Silent type coercion nearly wiped 9,291 cancellation records from retail dataset

A data engineer analyzing the UCI Online Retail dataset of 541,909 e-commerce transactions discovered that automatically parsing the InvoiceNo column as integers silently converted 9,291 cancellation records — prefixed with 'C' — into null values. The error went undetected because the first rows of the file parsed cleanly at 100%, masking the problem that only emerged when the full file was measured, revealing a ~98% parse rate. These cancelled orders represented 8.4% of total revenue, and losing them caused downstream figures to be quietly incorrect. The corrupted data also concealed a phantom bestseller — a single large order entered and fully cancelled within 12 minutes — which only the preserved cancellation rows could expose. The fix involved explicitly declaring identifier columns as string types, a one-line schema correction that saved all 9,288 cancellation records and produced a documented, replayable data-loading plan.