We actually got to a point where we had a couple of spare parts onsite (sticks of RAM, HD, etc) and so we repair immediately and then request the replacement. This was on a large HPC cluster so we had almost daily failures of some kind (most commonly we'd get a stick of RAM that would fail ECC checks repeatedly).