captured 2026-11-16T12:30:26Z --- heartbeat state RUNNING elapsed 0.426 down_now None {'allow': 195, 'anomalies': 0, 'cached_replies': 0, 'check_errors': 0, 'check_errors_expected': 0, 'checkpoints': 1, 'checks': 198, 'dag_submits': 0, 'deny': 3, 'divergences': 0, 'errors': 0, 'fault_command_failures': 0, 'faults': 1, 'max_unexpected_check_streak': 0, 'requests': 198, 'restarts': 1, 'unexpected_check_errors': 0, 'view_changes_seen': 8} --- fault / check lines 2026-11-16T12:24:58Z fault: killed cain-mr3-node-1 (atl) 2026-11-16T12:25:58Z fault: restarted cain-mr3-node-1 Reading this capture: - the soak's own fault injector killed cain-mr3-node-1 at 12:24:58Z and restarted it at 12:25:58Z (faults=1, restarts=1); - the run continued: requests 198, checks 198, divergences 0, errors 0; - unexpected_check_errors=0 and max_unexpected_check_streak=0 — the fault was not mistaken for a degradation; - check_errors_expected=0 as well, because the replica answered again before any check failed. The expected-error counter therefore did not need to increment on this fault; it will increment on a restart slow enough to fail a check (a restarted replica took ~140 s on 2026-11-16). The classification itself is covered by tests/test_soak_verdict.py.