Problem: sysbench --rate set well above cluster capacity overflows its internal event queue and hard-aborts (FATAL: event queue is full) in <15s instead of running a sustained high-latency window. Surprising for overload/latency testing.
Fix: document the semantics; recommend thread-bound runs (or capacity-relative rates) for over-capacity overload testing. (The --rate flag itself now exists.)
Found during the TiDB×sysbench load-test investigation (2026-07-21). See the run's final report for full context.
Problem: sysbench
--rateset well above cluster capacity overflows its internal event queue and hard-aborts (FATAL: event queue is full) in <15s instead of running a sustained high-latency window. Surprising for overload/latency testing.Fix: document the semantics; recommend thread-bound runs (or capacity-relative rates) for over-capacity overload testing. (The
--rateflag itself now exists.)Found during the TiDB×sysbench load-test investigation (2026-07-21). See the run's final report for full context.