Your last sentence hits the nail on the head. Even when storing petabytes of data, the odds on a freak hash collision are still many orders of magnitude longer than the odds on a hardware failure.
There's a statement that has been "pretty much permanently" on a whiteboard-covered wall of the computer lab at my college telling a joke about "the difference between a mathematician and an engineer", that goes through the math behind a specific type of prime number generator, calculates the likelihood that it might fail, and then claims the mathematician cares about that while the engineer knows that is orders of magnitude less likely than a guaranteed algorithm failing due to a cosmic ray hitting it in RAM and flipping one of its bits. ;P
Indeed... a quick search turns up someone from the zfs team indicating that a collision in zfs dedup (sha256) is 50 orders of magnitude less likely than an uncorrected hardware error. i shoukd have looked before posting.
