A 501(c)(3) non-profit organization Applied AI research for public benefit
CheAI Research Inc.

Increasing hazards in 8 of 22 hard-drive models and field failure rates above ratings for 20 of 21 in Backblaze data

We fitted hazard shapes to three years of Backblaze fleet data at each drive's own age and compared failure rates with ratings.

Abstract

On Backblaze Drive Stats data from Q3 2023 to Q2 2026, we tested pre-registered predictions on per-model hazard shapes and on observed against rated annualized failure rate (AFR). Each drive entered the risk set at its recorded power-on age, which handles drives installed before the window. Of 31 qualifying models, 22 had a usable age scale and a Weibull fit. The 95 % shape interval lay above 1 for 8 of them, below 1 for 5 and included 1 for 9, so wear-out was not shown to dominate. Early-life hazards decreased in 6 of 10 evaluable models, and 4 of 8 met the pre-registered bathtub definition. For 20 of 21 models with a determinable comparison, the whole interval on the observed AFR lay above the rating. The median observed-to-rated ratio over these 21 was 3.3. An independent implementation reproduced these tallies.

1. Introduction

Manufacturers rate hard drives by a mean time between failures (MTBF) or an AFR. For 22 of the 23 rated models here, the stated AFR is 0.35-0.44 %, an MTBF of 2.0-2.5 million hours. In field replacement logs for more than 100,000 drives, annual replacement rates were well above the datasheet values and rose steadily with age, with no bathtub [1]. Earlier work asked whether the rate settles after the first year [2] and criticised the bathtub curve [3].

Backblaze publishes a daily snapshot of every drive in its fleet [4]. Most drives were installed before the window opens, so the data are left-truncated, which biases the fitted shape unless each drive enters at its true age [5, 6]. We asked how per-model hazard shapes look once truncation is handled, and how many models differ from their ratings by more than the sampling uncertainty. Wear-out was not shown to dominate, early-life hazards often decreased, and the observed AFR exceeded the rating for 20 of 21 models.

2. Methods

The data are the 12 quarterly archives from Q3 2023 to Q2 2026, with 333,732,420 daily rows for 409,830 serial numbers. Each serial was split into episodes at gaps of more than 30 fleet days. An episode entered at the power-on age in self-monitoring, analysis and reporting technology (SMART) attribute 9 on its first row. Episodes with missing, implausible or inconsistent power-on hours were age-invalid. Models with over 20 % age-invalid drive-days received no shape result.

Thirty-one models had at least 100,000 drive-days and 20 failures. All were hard disk drives, labelled D01-D31 by nominal capacity. For each we fitted a Weibull likelihood with delayed entry and right censoring [7]. A shape β was increasing (wear-out) if its 95 % profile-likelihood interval lay above 1, decreasing (infant mortality) if below 1, constant if within 0.80-1.25, and otherwise undetermined. Episodes were also split at 1 year of age, and a segment needed 10 failures to be evaluable. A bathtub, here, means a decreasing early segment and an increasing late one.

The AFR is F/(DD/365), with F failures in DD drive-days, and has an exact Garwood interval for a constant rate [8]. Ratings were transcribed from manufacturer documents before any outcome was computed. A stated MTBF M was converted as r = 1 - exp(-8760 h/M), which assumes a constant hazard. With a stated service life, only age-valid exposure below it counted, and otherwise all exposure. A model was worse than rated if the lower interval bound exceeded the rating, and not determinable with fewer than 20 failures.

The rules and predictions were frozen in a pre-registration before any per-model outcome was computed [9]. On 50,000 synthetic drives, the delayed-entry fit recovered true shapes from 0.7 to 3.0 within 1.3 %, while ignoring truncation pushed them upwards by 27-117 %. A second implementation, written from the pre-registration alone, rebuilt all 31 models.

3. Results

3.1 Wear-out was not shown to dominate

Seven of the 31 models failed the age-validity rule and two gave no fit, which leaves 22 for the shape analysis. Eight had a shape interval above 1, five below 1, and nine included 1 (Figure 1). At 8 of 22 (36 %), the prediction of increasing hazards in most models did not hold. No model met the rule for a constant hazard.

Shape intervals for 22 drive models around a shape of 1
Figure 1. Overall Weibull shape for the 22 fitted models with 95 % profile-likelihood intervals. The dashed line is β = 1, and the grey band is the equivalence range 0.80-1.25.

3.2 Early-life hazards often decrease

Six of the ten models with an evaluable early segment were decreasing, with shapes from 0.21 (D23) to 0.55 (D16), so the prediction that infant mortality would not dominate did not hold. Of 20 evaluable late segments, nine were increasing and none was decreasing. Of the eight models with both segments, four (D11, D13, D19 and D23) met the bathtub definition (Figure 2), so the prediction of fewer than 25 % did not hold.

Early and late segment shape intervals by model
Figure 2. Weibull shape in the early (below 1 year, grey) and late (1 year or more, black) segments with 95 % intervals. Filled markers exclude 1, and open circles are undetermined. An asterisk marks the four bathtub models.

3.3 Observed failure rates exceed the ratings

Of the 23 rated models, 21 had at least 20 failures, and 20 of these had the whole 95 % interval above the rating (Figure 3). D31's AFR was 0.50 % (interval 0.31-0.74 %), consistent with its 0.35 % rating. No model was better than rated, so the rating prediction held. Over these 21 comparisons, the observed AFR was 1.4 to 27.0 times the rating, with a median of 3.3. On point estimates alone, D31 and D09 would be worse than rated and D04 better, although D04 and D09 have too few failures.

Observed failure rate intervals against the rating for 23 models
Figure 3. Observed AFR in the comparison exposure with the Garwood 95 % interval against the rated AFR (vertical bar) for the 23 rated models. Exposure is age-valid and below the stated service life, or all ages where none was stated (labelled). D04 had no failures, so its marker is the upper bound.

3.4 Reproduction and sensitivity

The independent implementation agreed on 2,047 of 2,051 comparisons, including every per-model regime and verdict. The four disagreements are the denominators of a secondary capacity-class prediction, whose numerators and verdict (not held) agree. At the Bonferroni level of 1 - 0.05/31 [10], the tallies fell to 4 of 22 increasing and 2 of 8 bathtubs, and the rated-AFR verdicts did not change. Early decreases fell to 2 of 10, which would reverse the early-segment verdict, but predictions are decided at the 95 % level.

4. Discussion

The late segments, never decreasing, agree with the rising replacement rates of the earlier field study [1]. Early life differs. Whether the rate settles after the first year [2] gets no single answer. In D19 and D23 an all-age shape below 1 hid a rising late hazard, so one shape should not be read as a failure mechanism [3].

The rating gap is consistent with the one reported earlier [1] and holds after sampling error, but the data cannot separate its causes. The failure flag and the failure definition behind the ratings come from different processes. The ratings assume a workload, up to 550 TB per year in most documents, that the snapshots do not record, and two of the 0.35 % ratings are labelled as projected. The verdicts describe this fleet's conditions, not whether a rating is met under the conditions it assumes.

5. Limitations

The age scale is SMART 9, which each drive reports itself. The validity rule removed 7 of 31 shape results, and for D07, 94.9 % of drive-days were age-invalid; a change of age source would need a new dated pre-registration. The data also do not show burn-in or deployment selection before entry, and truncation corrects only for survival to the entry age. Only 10 of 31 models had an evaluable early segment. The rated workload and temperature are not observed, and neither failure definition is documented in the sources used.

On the method side, the Garwood interval is exact only for a Poisson count with a constant rate, and the profile-likelihood interval is asymptotic under the fitted Weibull. Both assume independent failures and uninformative censoring and entry. These assumptions exclude shared operating conditions, removals that anticipate failure and differences between installation cohorts. The per-model intervals give no simultaneous coverage, and the Bonferroni results would reverse the early-segment verdict. The regimes classify the fitted Weibull, and D17 shows that one Weibull can misdescribe a model whose hazard falls early and rises late. The rule of at least 20 failures conditions the scope on the observed outcome.

The independent synthetic test found a coverage of 0.915 at β = 3 with 600 drives, near the late shapes of D17 (2.99) and the undetermined D30 (2.72). The D17 call is far from 1 and rests on 1,305 failures. Three recorded interpretations of the pre-registered rules concern exposure below the service life, capacity classes and a numerical tolerance. A fourth, the capacity-class denominator, was confirmed after the outcomes were known. Two supporting sensitivity analyses, a fit on time since entry and the AFR without age-invalid episodes, were not independently repeated. Independent Kaplan-Meier checks covered three models, and the rating transcription was not independently audited.

The fleet is one operator's workload and environment. The results make no claim about these models elsewhere, about models that did not qualify or about solid-state drives. A quarterly rerun is proposed but not scheduled, though later quarters would not change the verdicts reported here. Backblaze, Inc. publishes the data and asks users to cite it, to accept sole responsibility for their use, and not to sell the data. Manufacturer and model names are withheld.

6. Conclusions

Three years of Backblaze fleet data, analysed at each drive's own age, did not show the wear-out dominance that the pre-registration predicted. Early-life hazards fell in most models observed in their first year, and half of the models with both age segments evaluable showed a bathtub. In this fleet the observed annualized failure rate exceeded the rated value by its whole 95 % interval for all but one model with a determinable comparison.

References

  1. Schroeder B, Gibson GA. Understanding disk failure rates: What does an MTTF of 1,000,000 hours mean to you? ACM Transactions on Storage 3 (2007) 8. https://doi.org/10.1145/1288783.1288785
  2. Sun F-B, Zhang S. Does hard disk drive failure rate enter steady-state after one year? 2007 Proceedings Annual Reliability and Maintainability Symposium (2007) 356-361. https://doi.org/10.1109/RAMS.2007.328138
  3. Klutke G-A, Kiessler PC, Wortman MA. A critical look at the bathtub curve. IEEE Transactions on Reliability 52 (2003) 125-129. https://doi.org/10.1109/TR.2002.804492
  4. Backblaze, Inc. Backblaze Drive Stats, hard drive reliability test data. Quarterly archives Q3 2023 to Q2 2026, accessed 14 October 2026. https://www.backblaze.com/cloud-storage/resources/hard-drive-test-data
  5. Tsai W-Y, Jewell NP, Wang M-C. A note on the product-limit estimator under right censoring and left truncation. Biometrika 74 (1987) 883-886. https://doi.org/10.1093/biomet/74.4.883
  6. Hong Y, Meeker WQ, McCalley JD. Prediction of remaining life of power transformers based on left truncated and right censored lifetime data. The Annals of Applied Statistics 3 (2009) 857-879. https://doi.org/10.1214/00-AOAS231
  7. Lawless JF. Statistical Models and Methods for Lifetime Data, 2nd ed. John Wiley and Sons, Hoboken, NJ (2003). https://doi.org/10.1002/9781118033005
  8. Garwood F. Fiducial limits for the Poisson distribution. Biometrika 28 (1936) 437-442. https://doi.org/10.1093/biomet/28.3-4.437
  9. Nosek BA, Ebersole CR, DeHaven AC, Mellor DT. The preregistration revolution. Proceedings of the National Academy of Sciences 115 (2018) 2600-2606. https://doi.org/10.1073/pnas.1708274114
  10. Dunn OJ. Multiple comparisons among means. Journal of the American Statistical Association 56 (1961) 52-64. https://doi.org/10.1080/01621459.1961.10482090