A hard drive reliability check on 341,263 drives, from 4TB to past 20TB
Large cloud storage operators track their hard drives every day, recording which units keep running and which ones drop off the racks. Backblaze does this at scale, and its Q1 2026 report covers a fleet built for continuous use.
The analysis covered 341,263 hard drives, after boot drives and a small group of units that missed the reporting thresholds were set aside. The pool spans capacities from 4TB to past 20TB.
The quarterly annualized failure rate came to 1.24%. That figure rose from the prior quarter and settled under the level of a year earlier.
Across the whole fleet, the lifetime failure rate stands at 1.39%, a steadier long-run gauge of how the hardware holds up.
Newer, higher-capacity hardware carried much of the load. The company added more than ten thousand drives during the period, most of them holding more than 20TB each. This young pool posted an annualized failure rate of 0.85%, strong for freshly deployed equipment.
No new drive models joined the fleet this quarter, an outcome the company calls rare given that new models arrived in six of the past eight quarters. Several existing models ran clean, with the HGST HUH728080ALE600 and two others recording zero failures over the period.
Failure counts and failure rates tell separate stories. The Seagate ST16000NM000J logged a single failure and still posted a 3.61% annualized rate, a result driven by a pool that has dwindled to about 130 units. One drop from a small group moves the percentage a long way.
Several other models logged just one failure each across the quarter, among them the Toshiba MG09ACA16TE and a pair of Seagate drives in the 12TB to 14TB range. Their raw failure counts look calm, and their rates stay low because their populations remain large.
Backblaze counts failures with a C++ program that gathers SMART statistics at the end of each day. A drive present one day and absent the next gets logged as a failure, with a lookback window that reverses the call if the serial number returns within the quarter. The reported rates, author Stephanie Doyle wrote, come down to people “actively managing failure and risk.”
The method depends on a drive appearing in the pool the day before, which leaves day-one failures outside its reach. Some drives give out on their first day in a data center, and the counter can miss them. Day-one failures have probably gone under-recorded across the program’s history. Drive qualification before deployment keeps such cases rare, so the effect on the published figures stays small.
The complete dataset remains available for download.