Dataset · datasheet versus field · FAST 2007

Field disk replacements averaged 3.01% a year, 3.4 times the 0.88% a 1,000,000-hour MTTF implies

Schroeder and Gibson, FAST 2007, report a weighted average annual replacement rate (ARR) of 3.01% for drives under five years old across seven field data sets (more than 100,000 drives, SCSI, FC, and SATA), 3.4 times the 0.88% annual failure rate implied by a datasheet MTTF of 1,000,000 hours. ARR counts replacements, not confirmed failures. Replacement rates kept rising with age (7 to 10 times the datasheet rate in years 4 and 5 of HPC1, up to 30 times for 5 to 8 year old systems), and time between replacements in HPC1 fits a Weibull shape of 0.71, not an exponential model. For scale, Pinheiro et al., FAST 2007, report 1.7% to over 8.6% AFR in a Google fleet, and Backblaze reports 1.73% for Q2 2026 and 1.41% lifetime.

Download CSV18 rows · 7 columns · CSV

File /data/datasets/mttf-vs-field-replacement-rates.csv · JSON metadata · All datasets · Human-readable note on hesela.com

Method

The CSV transcribes figures printed in the text and tables of Schroeder and Gibson (FAST 2007), plus three cross-check figures from Pinheiro et al. (FAST 2007) and the Backblaze Q2 2026 Drive Stats post. It does not read chart values and it does not refit any distribution. The annual replacement rate (ARR) is drives replaced per year as a percent of drives in use. The datasheet AFR is power-on hours per year divided by MTTF.

Limits

ARR is drives replaced per year, not confirmed failures. The paper cites a vendor figure that 43% of returned disks show no problem.
ARR per data set cannot be recomputed from the paper's Table 1 because disk counts changed and logs hold other events.
Time-between-replacement, correlation, and Weibull figures are HPC1 only.
Several results appear only as charts in the paper and are not in this file.
Google age groups mix drive models. Backblaze figures are a different era and drive mix, and the opened post does not define a failure. Neither tests the 3.4 times ratio.

Table

18 rows, the same rows as the CSV. Value low and value high are the printed range ends, or a single printed value in the high column. Units are in the unit column. Sources are the pages opened on 2026-10-09.
Measure	Value low	Value high	Unit	Scope	Note	Citation
datasheet AFR at MTTF 1000000 hours	not stated	0.88	percent of drives per year	Vendor datasheet arithmetic: power-on hours per year divided by MTTF	Schroeder and Gibson print 0.88% for the 1000000 hour end of the range.	schroeder-usenix-html
datasheet AFR at MTTF 1500000 hours	not stated	0.58	percent of drives per year	Vendor datasheet arithmetic: power-on hours per year divided by MTTF	Schroeder and Gibson print 0.58% for the 1500000 hour end of the range.	schroeder-usenix-html
weighted average ARR for drives under five years	not stated	3.01	percent of drives per year	All seven data sets weighted by drive count	ARR is the annual replacement rate. A replacement is not a confirmed failure.	schroeder-usenix-html
weighted average ARR without COM3	not stated	2.86	percent of drives per year	Six data sets	The paper calls this 3.3 times 0.88%.	schroeder-usenix-html
range of ARR by data set and drive type	0.5	13.5	percent of drives per year	Drives within the five-year nominal life	One COM3 group deployed in 1998 had 24% and is excluded by the paper.	schroeder-usenix-html
HPC1 ARR over five years	not stated	3.4	percent of drives per year	HPC1: 765-node cluster collected August 2001 to May 2006	The paper says this is significantly higher than the datasheet AFR of 0.88%.	schroeder-usenix-html
weighted average ARR versus 0.88% datasheet AFR	not stated	3.4	times	All seven data sets	Observation 1 of the paper.	schroeder-usenix-html
replacement rate versus datasheet in years 4 and 5	7	10	times	HPC1 within nominal life	The datasheet MTTF implies the rate in the denominator.	schroeder-usenix-html
replacement rate versus datasheet for systems 5 to 8 years old	not stated	30	times	Older systems in the data sets	Printed as by as much as a factor of 30.	schroeder-usenix-html
correlation of replacement counts in consecutive weeks	not stated	0.72	correlation coefficient (unitless)	HPC1 only	Close to 0 would be expected for a Poisson process.	schroeder-usenix-html
correlation of replacement counts in consecutive months	not stated	0.79	correlation coefficient (unitless)	HPC1 only	Close to 0 would be expected for a Poisson process.	schroeder-usenix-html
Weibull shape of time between replacements	not stated	0.71	unitless	All HPC1 nodes	Shape under 1 means a decreasing hazard rate. Compute nodes 0.73. File system nodes 0.76.	schroeder-usenix-html
squared coefficient of variation of time between replacements	not stated	2.4	unitless	HPC1	An exponential distribution has a value of 1.	schroeder-usenix-html
probability of two drives failing within 1 hour versus exponential model	not stated	4	times	HPC1	Real data versus the fitted exponential model.	schroeder-usenix-html
probability of two drives failing within 10 hours versus exponential model	not stated	2	times	HPC1	Real data versus the fitted exponential model.	schroeder-usenix-html
Google AFR from first year to three-year-old group	1.7	8.6	percent of drives per year	Over 100000 ATA drives 5400 to 7200 rpm	Printed as 1.7% in the first year to over 8.6% for 3-year-old drives. Age groups mix drive models.	pinheiro-usenix-html
Backblaze quarterly AFR for Q2 2026	not stated	1.73	percent of drives per year	354415 drives and 31553350 drive days and 1498 failures from April to June 2026	Backblaze's own inclusion rules. The opened page does not define a failure.	backblaze-q2-2026
Backblaze lifetime AFR to June 30 2026	not stated	1.41	percent of drives per year	353041 drives and 558148607 drive days and 21609 failures	Models with more than 500 drives and more than 100000 lifetime drive days.	backblaze-q2-2026

Columns

measure (string): The quantity as the source names it.
value_low (number): Lower end of a printed range. Empty when the source prints a single value or an upper bound only.
value_high (number): Single printed value, or the upper end of a printed range, or the upper bound for 'as much as' wording.
unit (string): Unit of both value columns: percent of drives per year, times, or unitless.
scope (string): Population and window the value applies to.
note (string): What the value is not, or the source wording behind it.
citation_id (string): Id of the opened source in the citations list.

License

Small derived table of figures printed in Schroeder and Gibson (FAST 2007), Pinheiro et al. (FAST 2007), and the Backblaze Q2 2026 Drive Stats post, with credit. Not a Creative Commons license. The USENIX proceedings page says rights to individual papers remain with the author or the author's employer. This file is not a copy of the papers, not the replacement logs, and not the Backblaze dataset. The Backblaze post asks that its raw Drive Stats data be cited and not resold, which this table does not redistribute.

Sources

Disk failures in the real world: What does an MTTF of 1,000,000 hours mean to you?, Schroeder and Gibson, FAST 2007 (USENIX HTML proceedings) (accessed 2026-10-09)
Disk Failures in the Real World, USENIX FAST 2007 abstract page (accessed 2026-10-09)
Failure Trends in a Large Disk Drive Population, Pinheiro, Weber, and Barroso, FAST 2007 (USENIX HTML proceedings) (accessed 2026-10-09)
Backblaze Drive Stats for Q2 2026, published September 29, 2026 (accessed 2026-10-09)