Citadel Software Engineer Interview Experience — A Maddeningly Vague Data Sampling Question

Citadel is out of their minds.

The first half was fine — a normal self-introduction, project walkthrough, basic behavioral questions.

Then the coding part started and that's where the craziness began:

The interviewer was mumbling and couldn't speak clearly, and gave me a purely verbal problem: "we have millions of data points that need to be labeled," and told me to find the useful data. I kept following up, and he finally said "get 200 out of 1 million data points." I asked "what counts as useful data," and the interviewer wouldn't say. I asked how he wanted me to sample, and he wouldn't say that either.

OK fine, I proposed random sampling — he said that works but isn't optimal. I brought up least recently used, most frequently used, average, median, p99, p75, that kind of thing — he said it can only be called once. I followed up again asking "do I need to assume an object and then come up with several useful features" — he said he didn't know what features even meant.

Can someone please tell me what on earth this company, and this question, actually WANT! TO! DO!!!!!!