Sustainable Systems Laboratory good data & systems that last

Technical Report SSL-TR-2026-001

Drive Failure and Lifetime Simulation in the Backblaze Drive Statistics Corpus, 2013–2026

Preprint v0.1 · October 17, 2026 · © 2026 The authors.

Abstract

Storage simulators estimate the chance of data loss from assumptions about drive lifetimes, differences between drive models, and possible dependencies between failures. Common lifetime models include a simple exponential distribution, a single Weibull curve, and a fixed three-component bathtub distribution. We test these assumptions against Backblaze's daily records from April 2013 to June 2026. From this 13-year trove of data, we identify 504,624 data drives with a usable power-on age, drawn from 99 drive models (after reconciling equivalent names), which amounts to 2.0 million drive-years. The records include 31,881 failures: removals because a drive physically failed or was judged likely to fail. Accounting for age and drives already in service when observation begins, we fit models to earlier records and test their predictions against nine successive later years. We also forecast the drives in service at each forecast date.

We find that accounting for differences between drive models helps more than any fine distinctions among the leading lifetime distributions. We assess forecasts with a log score, which rewards assigning greater probability to the failure counts that actually occur. Adding drive-model differences and attributes to a two-Weibull baseline improves this score about 4.3 times as much as replacing a single Weibull with two competing Weibulls while retaining the same drive-model treatment.

Fleet failure risk falls in early life and rises at older ages, giving the fitted age curve the familiar bathtub shape.¹ The operational failure records do not identify the underlying physical mechanisms. Two competing Weibull hazard components describe this pattern well, but several alternatives predict similarly. The fixed bathtub benchmark predicts 2.4 times the observed failures. Allowing unusually good or bad years matters too: on 169 well-observed combinations of drive model and year, selected using earlier records, forecast ranges intended to contain 95% of annual failure counts contained only 44% when annual variation was omitted, but 92% when we allowed both fleet-wide changes and unusually good or bad years for individual drive models. Useful average failure forecasts can nonetheless substantially understate uncertainty.

For an existing fleet, current condition adds information beyond model and age. A SMART (Self-Monitoring, Analysis and Reporting Technology) warning improves failure-risk ranking: the one-year area under the receiver operating characteristic curve (AUC) rises from 0.64 to 0.73. Gradient-boosted trees using monthly SMART records perform best on our measure of individual failure probabilities, with AUC 0.74 at one year and 0.83 at 90 days, yet predict about 19% and 22% too many failures, respectively. In other words, identifying individual high-risk drives, and forecasting replacement demand, require separate checks.

Long-term survival is harder to predict. Drives age as calendar time advances, and newer drive models enter a changing fleet, making these effects difficult to separate. Different fitted accounts of aging and calendar time imply fifteen-year survival probabilities up to 18 percentage points apart; assumptions about future calendar trends produce differences of up to 27 points. It is therefore critical that any simulations that depend on long lifetimes should test their sensitivity to these assumptions. We supply implementations that generate simulated lifetimes, represent shared variation, and model changes between SMART warning states, failure, and retirement. Allowing failure risk to change from year to year is computationally inexpensive: in our benchmark, generating a drive's lifetime takes about 154 nanoseconds of computation for each year of simulated service.

¹ The usual intuition is that drives with latent defects would tend to fail early. Those that remain in service without failing early could therefore be assumed to face a period of lower failure risk before wear-out becomes more important as they reach older ages.

Cite as

Ahmed Amer and Darrell D. E. Long. Drive Failure and Lifetime Simulation in the Backblaze Drive Statistics Corpus, 2013–2026. SSL Technical Report Series, SSL-TR-2026-001. Sustainable Systems Laboratory, Santa Clara University, Draft of 2026-10-17. Preprint v0.1. sustainable-systems-lab.org/reports/SSL-TR-2026-001/

Version history