Drive reliability

Drive policy

Run a SMART extended self-test on every used drive before it enters service.
Run a ZFS scrub after every resilver.
Record power-on hours, SMART attributes 5, 187, 188, 197, and 198, temperature, and the self-test result with its LifeTime hour. Store them in the hardware inventory, keyed by the last 4 serial characters.
Read self-test results with smartctl -l selftest. The -c summary reports “completed without error” even when no test ran. Confirm the LifeTime hour advanced.
A missing SMART attribute is not zero. State when a model does not implement one.
Record measurements, not predictions. Do not call a drive “dying” or “end of life.”
Redundancy absorbs drive failure. Replacing a healthy drive by age is not a protection strategy. See data protection classes.

Backblaze Q1 2026: 1.24% quarterly annualized failure rate (AFR) and 1.39% lifetime AFR across its drive population.
Failure-rate peaks moved later: 3 years 3 months for the 2013 cohort, 7 years 9 months for the 2021 cohort. Age alone is not a defect.
SMART attributes correlate with failure. They do not give a calibrated failure probability for one drive.
Drives bought and run together share age, so they fail together more often than independent drives. That matters during a rebuild.
A non-zero attribute 197 raises failure probability. It does not prove the drive is dead. Retest it standalone before any disposal decision.