ML job throughput considerations