Field disk replacements averaged 3.01% a year, 3.4 times the 0.88% a 1,000,000-hour MTTF implies

The headline gap. The weighted average ARR across the data sets, for drives within their five-year nominal life, was 3.01%. The paper calls this 3.4 times the 0.88% that a 1,000,000-hour MTTF implies. Without the COM3 data, which had the highest rates, the average was still 2.86%, or 3.3 times 0.88%. The observed ARRs by data set and drive type ranged from 0.5% to 13.5%, up to a factor of 15 above the datasheet AFR. HPC1, which covers almost exactly the five-year nominal life, had an ARR of 3.4%. One group of COM3 drives deployed in 1998 had an ARR of 24%, but those were more than seven years old and outside the nominal life.

The gap grows with age. Replacement rates were higher than the datasheet suggested in every year except year 1, and the HPC1 file system nodes had no replacements in the first 12 months. In HPC1's second year the rates were 20% above expectation for the file system nodes and a factor of two for the compute nodes. In years 4 and 5 they were 7 to 10 times what the datasheet MTTF implies. For systems 5 to 8 years old, datasheet MTTFs understated replacement rates by as much as a factor of 30. The paper finds no steady state after year 1 and no strong infant mortality. It sees a steady rise that it calls an early onset of wear-out, and it recommends that reliability standards include wear-out. HPC1 replacement rates nearly doubled from year 1 to year 2, and from year 2 to year 3.

The independence model does not fit. A Poisson process, with independent failures and exponentially distributed gaps, is the usual assumption behind RAID loss estimates. In HPC1 the correlation coefficient between replacement counts in consecutive weeks was 0.72, and between consecutive months 0.79. The expected number of replacements in a week varied by a factor of 9 depending on whether the preceding week was in the lowest or the highest third. The time between replacements was fit best by a Weibull distribution with a shape of 0.71 over all HPC1 nodes (0.73 for compute nodes, 0.76 for file system nodes), and the exponential model was rejected at the 0.05 level. The squared coefficient of variation was 2.4, against 1 for an exponential distribution. A shape under 1 means a decreasing hazard rate: right after a replacement the expected wait for the next one was about 4 days, but after 10 days without a replacement it had grown to 10 days, and after 20 days to 15 days. For a rebuild window of a few hours, the paper reports that two drives failing within one hour was four times more likely in the real data than under the exponential model, and within 10 hours two times more likely.

Where a drive sits among the parts that get replaced. In HPC1 the hard drive was 30.6% of all hardware replacements, in COM1 18.1%, and in COM2 49.1%. Among HPC1 node outages attributed to hardware, drives were 16% of outage records, behind CPU at 44% and memory at 29%, but about 90% of the drive problems led to a drive replacement, a costlier repair than the simple fix that cleared most CPU and memory events.

Two other sources agree on the level, not on the ratio. The Google paper (Pinheiro, Weber, and Barroso, FAST 2007) reports annualized failure rates from 1.7% in the first year to over 8.6% in the three-year-old population for more than 100,000 consumer-grade ATA drives, and says vendors often quote yearly rates below 2%. The Backblaze Q2 2026 report gives 1.73% for the quarter (1,498 failures over 31,553,350 drive days, 354,415 drives) and 1.41% over the lifetime of its fleet (21,609 failures over 558,148,607 drive days, 353,041 drives). As my own arithmetic, 1.41% is about 1.6 times 0.88%. That is a different era and drive mix, counted under Backblaze's own rules, so it is not a replication of the 3.4 times figure. The Google paper warns that its age groups mix drive models, so its age pattern is not a pure aging effect; the age finding here therefore rests on one study.