Estimate annual hard drive failure probability from AFR and fleet size.
Annualized Failure Rate (AFR, as published in studies like Backblaze's drive stats) gives a one-year failure probability per drive; compounding it over multiple years uses P(failure) = 1 − (1 − AFR)^years, assuming a constant hazard rate (a simplification — real drives follow more of a bathtub curve, with elevated infant-mortality and end-of-life failure rates). When you only have an MTTF/MTBF rating instead, AFR is derived from the exponential reliability model AFR = 1 − e^(−hours_per_year / MTTF). For a fleet of drives, the probability that at least one fails is much higher than any single drive's rate — this is the core justification for RAID redundancy at scale.
Multi-year failure probability
P(failure over N years) = 1 − (1 − AFR)^N
AFR from MTTF
AFR = 1 − e^(−8760 / MTTF_hours)
Because it's the probability that at least one of many independent drives fails, which compounds quickly: with a 1.5% AFR, a single drive has a 98.5% chance of surviving a year, but 100 drives together have only 0.985^100 ≈ 22% chance that all 100 survive — meaning a 78% chance at least one fails.
No — MTBF/MTTF is a statistical rate derived from failures across a large population under the assumption of a constant (exponential) failure rate; it does not mean an individual drive is expected to run for that many hours. A 1.2-million-hour MTTF drive can still fail in year one, and virtually none are expected to run continuously for 137 years.
No — this uses a constant hazard rate model, which is a simplification. Real drive failure rates typically follow a bathtub curve: higher in the first few months (infant mortality/manufacturing defects), lower during a stable middle period, then rising again as drives age past 3-5 years — AFR-based estimates are most accurate for that stable middle period.