Your Ray Data Pipeline Works at 10K Samples. Here's Why It Crashes at 1M.

There's a moment every ML infrastructure engineer knows: the evaluation pipeline that worked...