What fails most often: unreadable data on disks, uncorrectable reads on SSDs, and bus errors on the link

The direct answer to what fails most often is unreadable data, not a drive that stops. On hard disks, 3.45% of 1.53 million disks (53,820) developed at least one latent sector error over 32 months, which is an error found only when the sector is read; nearline disks were affected at 8.5% and enterprise disks at 1.9%. Google's flash paper restates the disk figure as 3.5% over 32 months, so the number is repeated by an independent group, but it is the same underlying dataset and not a second measurement. On SSDs, the two public field studies measure read failures on different populations and agree on direction, not on size: in Google's drives (Table 2 of the flash paper), 20.3–90.5% had at least one uncorrectable error and 19.1–62.7% at least one final read error (a read that still fails after retries) in the first four years (three for the eMLC models), and 4.2–34.1% of Facebook's SSDs per platform had an uncorrectable error.

Errors cluster rather than spread evenly. In the disk study, the median error disk had 3 errors, the mode was 1 error (30% of error disks), and 0.2% of error disks had more than 1000. In the Facebook SSD data, the 10% of SSDs with the most errors held over 80% of all uncorrectable errors on every platform. In Google's flash data, a drive with an uncorrectable error had nearly a 30% chance of another in the next month, against about 2% in a random month, and in the MLC models half of the drives that reached two bad blocks went on to about 200 or more. That is the practical reason a first error matters more than a count.

On the link, the evidence is thinner. The Google disk study reports CRC errors on about 2% of its disks and says they point to cables and connectors more than to the drive. The Linux libata documentation agrees on the classification: an ICRC bit means corruption during transfer and is handled as an ATA bus error, not a device error, and a SCSI sense of HARDWARE ERROR with ASC/ASCQ 47h/00h (parity error) is also treated as a bus error. By contrast the UNC bit is a media error reported only after the drive's retries failed. Neither source gives a fleet-wide link failure rate, so none is stated here.

The counters miss failures. Google found over 56% of failed drives had no count on scan errors, reallocation, offline reallocation or probational count. Backblaze reported that 76.7% of failed drives had at least one of five SMART attributes (5, 187, 188, 197, 198) above zero, leaving 23.3% with none, while 4.2% of operational drives also had a non-zero raw value. Those two numbers use different attribute sets and cannot be averaged. For SSDs, Google found the raw bit error rate is not a good predictor of uncorrectable errors, so a rising corrected-error count alone does not say a read failure is coming.