Google is a prime user of commodity grid supercomputing. As such, they are in a perfect position to release a paper on disk reliability, with hundreds of thousands of data points. Very interesting results, including the failure of SMART to predict failures, and low correlation with usage and temperature.
This will be important as grid supercomputing becomes the preferred way to manage compute resources.
Update: Slashdot points to another storage paper presented at FAST ’07 confirming some of the points in the Google paper, and invalidating the manufacturers’ MTTF estimates.