Machine Learning System Design Interview, issue 32, May 20, 2026
The Distributed Pandas Trap
Senior ML Platform Engineer interview at OpenAI, and the interviewer asks:
“A data scientist hands you a feature engineering script built natively in Pandas that runs perfectly on their local 16GB laptop. They want to move it directly to production to process a 5TB daily log stream. How do you containerize and scale it?”
The hidden memory amplification trap that turns local feature scripts into cloud bill nightmares, and why true scalability requires changing the data layout, not vertical scaling.
The full answer, with the mechanism and the arithmetic, is for paid subscribers on Substack.