Performance testing checks whether a data-pipeline task or model meets defined execution-time, resource-use, or evaluation conditions. Treat performance as context-dependent: accuracy alone may not represent what matters. section 5.4.3 section 8.6
Turn performance expectations into explicit checks:
- Set absolute limits, such as completing a task in fewer than three seconds, or relative limits, such as staying within two standard deviations of normal runtime. section 5.4.3
- Measure nonfunctional properties including time, memory, cost, and invocation count alongside functional results. section 5.4.3
- Choose model metrics deliberately, potentially combining precision, recall, F1 score, and other measures rather than relying on accuracy. section 8.6
- Record privacy and security constraints, testing trade-offs, and the reasoning behind model comparisons or selections. section 8.6
Pipeline runtime check
A pipeline step passes when it completes in under three seconds; a broader test can compare runtime with its normal distribution. Failures should feed monitoring, logging, notifications, and alerts. section 5.4.3
