Did the author actually verify this with strace (or the mac/windows equivalent)?
It sounds like he guessed this based on I/O activity of the process. It could be enough to hash the beginning of the files, and compare the rest if a match is found in the database.
Not really, one could get a unique enough hash by reading the first lets say 10,000 bytes of each file, and it would be faster than hashing the whole file
I wrote something that was hashing audiobook files that was taking forever, so I tried using the first N bytes (likely much more than 10kB), but soon found that for any given audiobook, each chapter's MP3 had a large identical header on the front end - I imagine that it was a cover image embedded in the metadata.
I think in the end I just started taking the data from the end of the file, but if you're going with subsets, it's probably better to use a pseudo-randomly selected subset rather than a sequential subset. It doesn't have to be a different pseudo-random subset for each file, but I imagine there's an ideal noise profile in the sampling (maybe white noise is best).
Of course you are correct (not sure why my comment was downvoted) but in the context of having a unique enough hash TAKEN QUICKLY 99.999% of time in set of millions of files its good enough, if one needs better hashing they can hash the whole file but this is quite heavy on large files and pointless if there is no need for it by the application