[Google] SWE-ML (L4) loop: the ML system design round was the differentiator

Google ML SWE (L4): the DSA coding rounds were standard-ish, but the ML system design round is where it's won or lost.

They wanted a clean retrieval → ranking split, how I'd keep features fresh, and how I'd measure quality online vs offline. Got asked to defend NDCG vs recall@k for the specific use case, not just name metrics but justify the choice.

Coming in from pure backend, that round needed real, separate prep. Don't assume general system-design reps cover it.