Skip to main content
Base rates for reasoning about HDD risk, and the recording standard that depends on them. The purpose of this page is to replace adjectives with numbers.

​
The rule

No drive record calls a disk “dying,” “worn,” “end-of-life,” or “on borrowed time.” Those are predictions written as observations. Record the measured values, cite a base rate from this page, and let the reader do the arithmetic. This is not stylistic. Age-based intuition about disk risk is often wrong — the data below show failure rates that don’t track age the way most people expect.

​
Source data

Backblaze Drive Stats Q1 2026 — 341,263 drives, 30,203,180 drive-days in the quarter, 529,968,464 lifetime. AFR (annualized failure rate) is the share of drives expected to fail over a year of continuous service.

​
Failure vs. age

The finding that matters most, from Are Hard Drives Getting Better? Let’s Revisit the Bathtub Curve: The age at which the failure rate spikes moved later by roughly two years between those cohorts. Backblaze’s own conclusion is that the classic bathtub curve no longer describes their population: “the neat story of early failures, calm middle age, and gentle decline no longer fits the world our drives inhabit.” Consequence. A 6–7.5 year old drive is not self-evidently near failure. It sits inside the band where a modern cohort still shows a low per-drive AFR, at or before the age where the 2021 cohort’s rate peaked. Age alone is not a defect and is not recorded as one.

​
What actually predicts failure

Backblaze and Google independently identified a small set of SMART attributes that correlate with impending failure. Backblaze uses five, chosen because they are consistent across manufacturers: Google’s analysis found 5, 187, 197, and 198 consistently correlated, and reported that after a drive’s first scan error — SMART 187 going non-zero — it is 39× more likely to fail within 60 days than a drive with none. Two limits, because they change how strongly these may be cited:
  • Both analyses are univariate, one attribute at a time. They establish correlation, not a calibrated per-drive probability.
  • Not every model implements every attribute. A missing attribute is not a reading of zero. Where one is absent, say so and substitute the evidence that does exist — normally a completed extended self-test.

​
Recording standard

  1. Record measurements, not prognosis. Power-on hours, the five predictive attributes, self-test outcomes with their LifeTime hour, temperature.
  2. Read self-test results from smartctl -l selftest, never the -c summary. -c reports “completed without error” even when no test has run; confirm the LifeTime hour advanced.
  3. State absence explicitly. “Attribute 197 not implemented on this model” — never silently imply zero.
  4. Cite a base rate rather than an adjective. Record the measured pattern — power-on hours, each predictive attribute’s value, the self-test result — instead of a subjective label. A full attribute readout is checkable; “high-hours refurb nearing end of life” is not.
  5. Redundancy is the control, not drive selection. A non-zero population AFR means any drive can fail at any age — that’s what in-pool redundancy and off-site copies absorb, not preemptive replacement. See ZFS backup and replication for how those layers combine.

​
Worked examples (not published data)

The statistics above are population-level base rates from Backblaze and Google. Applying the recording standard to one drive is a different exercise. The two cases below are invented to illustrate the reasoning — they are not measurements, and not part of either dataset. Example — keep in service. A drive several years into service reads zero on every predictive attribute it implements, and completes a full extended self-test without error. Nothing measured on it indicates a defect, regardless of its age. Example — pull from the pool. A drive that accumulates Current Pending Sectors — even with Reallocated Sectors and Offline Uncorrectable Sectors still at zero — has crossed one of the five predictive attributes. That’s evidence-backed grounds to remove a disk from a pool that can’t absorb another failure, independent of the drive’s age. In this second example, a non-zero pending-sector count raises failure probability materially, but it doesn’t mean the drive has failed: zero reallocated and zero offline-uncorrectable sectors means the platters haven’t lost data. Retest a pulled drive standalone before any disposal decision. Correlated age is a pool property, not a drive property. Disks bought and deployed together tend to fail less independently than the base rate assumes — which matters most during a rebuild, when the surviving disks take their heaviest read load. That’s an argument for rebuild-window planning and verified restores, not for treating any individual disk as suspect.