benchRun() calls .detach() on the worker threads it spawns. This means that despite calling _brState.waitForState(BenchRunState::BRS_FINISHED) to wait for the worker threads to not be considered "active", it doesn't wait for the worker threads to actually be destructed. There is then a race in mongoebench where the destructor for the thread_local ServiceContext::UniqueClient currentClient may not have run by the point embedded::shutdown() is called. Having benchRun() explicitly .join() the worker threads it spawns would fix this race.

void BenchRunWorker::start() {
    stdx::thread([this] { run(); }).detach();
}