I agree, and I try to make clear that I really want to understand how the candidate would approach it -- if they genuinely would try to find a perfect hash and

I agree, and I try to make clear that I really want to understand how the candidate would approach it -- if they genuinely would try to find a perfect hash and keep a hash -> offset map in memory, then mmap a big data file, that is fine. I will ask how they will find the perfect hash, etc. If you can describe well the approach you take, then this is fine. Ofttimes, though, candidates have started designing a general purpose on-disk hash database. That is fine and dandy, and we always need a better database, but it is pretty orthogonal to the immediate problem. Actually, ofttimes engineers have implemented general purpose on-disk hash databases as part of j. random project, so that path certainly does occur on real projects :-)