Default Kubernetes configuration can be suboptimal for ML workloads, resulting in wasted human and computer time. We show how a round of troubleshooting and a two-line change in the code yields a nine-fold speedup in the total execution time.