You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Load mode grows from constant-vus to four executors: ramping-vus (stages with graceful ramp-down), and the open-model constant-arrival-rate and ramping-arrival-rate, which start iterations on schedule whatever the response time and report dropped iterations once max_vus are busy. A load: entry can add think_time, thresholds (e.g. p95: <300ms, errors: <1%, per step under steps:), checked at the end and failing the run with exit code 99, and abort rules evaluated every metrics frame that stop a run early with exit code 108. --report frames:<url> streams every metrics frame, with per-step HDR histograms, and the final report to an HTTP endpoint (bearer token from DEVTOOLS_FRAMES_TOKEN, retried in the background, and the run stops if the endpoint is unreachable for 60s); --vus-scale/--rate-scale split one profile across machines, --load-file reads scenarios from a separate file with the same schema, and the summary now shows the request count it is based on. (c5002b3d)