Whenever you're ready, walk me through how you'd approach this.
Describe a time you had to troubleshoot a performance bottleneck in a distributed machine learning pipeline.
Explain how you localized the bottleneck, distinguished infrastructure issues from data-loading or model-computation issues, and validated the improvement. Include the profiling tools, measurements, experiments, corrective action, and safeguards you used to ensure that training correctness and model quality were not degraded.