LEARN. PRACTICE. GROW.Concept and mechanism
A performance target should describe operation, population, load, and metric. A mean can conceal a small share of very slow responses; a percentile also needs suitable sampling and segmentation. In a portal with fast queries and heavy exports, the operation mix influences results. Do not extrapolate a query-only experiment to daily cutoff with concurrent exports. Record version, data volume, cache preparation, and environment characteristics. Compare what the generator intended to send with what it actually started and completed. Low latency with many omitted requests may mean the desired load never reached the system, limiting the capacity conclusion.
Guided application
In a closed model, each virtual user starts the next iteration after completing the previous one. If the service slows, arrival rate may fall. An open model attempts to maintain arrivals independently of duration but remains limited by generator resources. In k6, dropped_iterations may indicate insufficient available virtual users or increasing work duration; it does not identify the cause by itself. Use generator and system metrics to separate hypotheses. Checks verify individual observations, while thresholds define criteria that can make the run exit with failure. Keep these criteria connected to the objective, including error rate and delivered load, to avoid an apparently positive report that misses the intended usage condition.
Planned and delivered load belong in the same analysis.
Common pitfalls
Mean treated as tail; configured load treated as delivered; check treated as exit status.
Related topics: Technical risk and operational evidence · White-box logical coverage · Static and dynamic analysis
Validate generator and profile before concluding capacity.
Reference: Grafana k6 open and closed workload models · CTAL-TTA v4.0 (2021)