Stories about floating-point arithmetic
1 related stories
Bounded, Indeterminate, or a Bug: A Condition-Aware Oracle for Differential Testing of SQL Aggregates
AI InsightTraditional differential testing treats cross-engine result discrepancies as bugs, but floating-point aggregates are inherently non-associative, causing frequent false positives. This research uses exact rational values as ground truth and proposes a condition-aware classification that labels discrepancies as exact, bounded, or indeterminate, shifting the basis from the query to the engine's algorithmic properties. This suggests that reliability of differential testing no longer depends on empirical epsilon but on a mathematical characterization of error propagation.Key TakeawayDifferential testing of SQL aggregates is shifting from 'discrepancy means bug' to condition-aware classification based on error bounds.Why It MattersExisting differential testing either relies on epsilon patches or avoids floating-point scenarios entirely for aggregates. This oracle offers an exact mathematical baseline to automatically distinguish real defects from legitimate errors, potentially reducing false positives in database testing and enabling more automated verification of SQL aggregate implementations. It also informs differential testing methodology for other numerical-intensive systems such as ML inference.Who's Affected- Database DevelopersCan integrate this oracle into test suites to automatically filter legitimate floating-point errors, reducing manual review costs.
- Differential Testing ToolsGain a theoretically grounded test criterion to replace empirical epsilon, improving test trustworthiness.
- ResearchersIf error bounds extend to more algorithms or operators, it may yield new test coverage strategies.
What's NextNext, watch whether this oracle gets integrated into mainstream database testing frameworks such as SQLancer, and whether its error-bound formula generalizes to non-aggregate operators. If tools emerge based on this method or broader empirical evaluations appear, its practical value will be confirmed.Importance 66/100