But Fermi problems are quite relevant in computing work. Here is a question I was asked recently:
"How many users do you think will ever concurrently use a certain subset of our website?"
Actually, the question was phrased as "how would you architect Feature X?" This naturally led to "do we need a distributed approach?" And of course, the answer to that depends on the answers to the Fermi question and "how many users can a single box handle".
I've never been asked a to solve a Fermi problem in an interview, but I would most likely require that the interviewer give me the data I need. In the golf ball example, I would ask for diameter of the golf ball, dimensions of a bus, etc. Otherwise they're just asking me to prove that I have seen a golf ball up close, that I've been riding on double-decker buses and that I'm decent at measuring physical objects with my eyes.
Or you could guesstimate a diameter of a golf ball and then tell them how much the answer could change as a function of your error, rather than demanding perfect information in an imperfect world.
The goal is not to get a precise answer, the goal is to see how you think.
Sorry, I didn't mean to imply I wanted precise information. Their guesstimate is going to be better than mine, which is why I would ask them for that info. I'm just not comfortable with pulling data out of my lower dorsal region. And I guess that's the part of how I think.
Another reason I like Fermi problems, is that the kind of people who I want work with (read: curious-minded tend-to-be-geeky programmers) tend to love them, and solve them for fun. I can't count the number of times we sat over lunch, arguing weird questions like how many planes it would take to move a certain river, or how many people it would take to eat a rhinocerous.
If someone is put off by these kinds of questions, I would question whether they fit with the kind of culture I think a group of smart people have.
"How many users do you think will ever concurrently use a certain subset of our website?"
Actually, the question was phrased as "how would you architect Feature X?" This naturally led to "do we need a distributed approach?" And of course, the answer to that depends on the answers to the Fermi question and "how many users can a single box handle".