Reading Backblaze Drive Stats before you buy

A table with 31 drive models and a very confident percentage column looks like a shopping guide. It is a lab report from somebody else's building. This guide shows how to read it without buying the wrong conclusion.

The short answer

Every quarter, Backblaze publishes how many of its own data-center drives failed, per model, as an annualized failure rate (AFR). The raw data is public, which makes it one of the few large open sources on how drives age [1][2]. It cannot tell you how a particular disk will behave in your NAS, and for many models the sample is too small to conclude anything firm. Use it to ask better questions about a model: how many drives, how old, how many drive days, how many failures. Then rely on warranty, SMART monitoring and a tested backup. We do not rank or recommend individual models in this guide.

How it works

1. What the data is. Since 2013, Backblaze has published the failure statistics of the drives in its data centers. Every day it takes a snapshot of each operational drive, including the S.M.A.R.T. values the drive reports (SMART is the drive's built-in self-monitoring). One drive on one day is one row, and the daily files are free to download [2]. The newest report on 24 Oct 2026 covers Q2 2026 (1 Apr to 30 Jun 2026) and appeared on 29 Sep 2026. Backblaze monitored 359,101 data drives, removed 3,881 boot drives and 705 drives that did not meet the criteria, and analyzed 354,415 drives with 31,553,350 drive days and 1,498 failures, which is an AFR of 1.73% [1]. Backblaze calls that the highest quarterly rate in quite a while [1]. For comparison, the page lists 1.41% for the lifetime and 1.36% for the year 2025 [2]. We did not find a Q3 2026 report on that date.
2. How the AFR is computed. Drives come and go in a data center, so a simple "failures divided by drives" would be wrong. Backblaze adds up drive days (one drive running for one day) and annualizes: AFR = failures / (drive days / 365) × 100 [4]. Backblaze explains that a drive-count method would assume a stable number of drives, which its fleet does not have [4]. You can check any row. In Q2 2026, one 12 TB Seagate row shows 1,671,057 drive days and 145 failures: 145 / (1,671,057 / 365) × 100 is 3.17, the published value [1]. We chose that row for the arithmetic only. It is not a verdict.
3. What counts as a failure. Backblaze names two kinds. Reactive failures are drives that stop working or no longer talk to the system. Proactive failures are drives that report errors, confirmed through their SMART values, and are pulled before they die completely [4]. So a failure in the table can be a drive that Backblaze replaced early. The Q1 2026 report adds how the program finds failures: it logs a drive as failed when it was present the day before and is missing today, with a lookback period for drives that reappear. Because of that, Backblaze says day-one failures are probably under-reported, since drives have already passed qualification before they enter the data [3].
4. Who is in the table. Models enter a table only above a minimum: quarterly more than 100 drives and more than 10,000 drive days, annual more than 250 and 50,000, lifetime more than 500 and 100,000 [1]. Smaller models are left out of the charts but remain in the public dataset [1]. In the lifetime table of Q2 2026, four models are missing because they fall short of the 500-drive minimum, two of them with zero failures in the quarterly table [1]. A missing model is not a bad model. It is a model Backblaze has too little data on. If the model number on your drive's label does not appear, the table says nothing about it.
5. Sample size decides how far you can trust a percentage. Ten thousand drive days is only about 27 drive-years. At that size one failure moves the AFR by about 3.65 points. Our rule of thumb: one failure changes the AFR by 365 / drive days × 100 points. For a Q2 2026 row with 86,092 drive days that is 0.42 points. For the row with 1,671,057 drive days it is 0.02 points (our rounded calculation). Backblaze itself flags this: two Seagate models had AFRs of 9.33% and 8.26% with only 965 and 1,235 drives, 22 and 25 failures [1]. Backblaze counts three outliers above its quartile-based threshold of 6.95% and gives age and population size as reasons [1]. According to Backblaze, ten of 31 models were above 3.0%, which it calls high compared with other quarters and "something to keep an eye on" [1].
6. Age matters as much as the model. Every row has an average age in months. In Q2 2026, Backblaze points to a nearly seven-year-old HGST model and a Seagate model of nearly 8.5 years among its outliers [1]. Young drives have a short history: the five youngest rows in that table average between 3.4 and 12.2 months [1]. A low AFR on a young drive is not a statement about year five. Backblaze warns that failure rates follow the bathtub curve "unless it doesn't", and that some drives are great and then "hit the wall" [6]. A large field study of disk replacements found rates that grew with age from the second year, not a flat middle phase [10]. Compare the quarterly table with the lifetime table of the same model and read both ages.
7. Confidence intervals, as far as Backblaze gives them. A confidence interval gives the range in which the true rate probably lies. Backblaze has described its own use of a 95% interval, the gap between the low and high bound: it likes the gap to be 0.5% or less, and a larger gap means it needs more data or the data is inconsistent [5]. In 2022 it wrote that it would need at least 500,000 drive days in a quarter "to get to a 95% confidence interval" for models with zero failures and few drive days [4]. The Q2 2026 tables, which we read on 24 Oct 2026, list drive count, average age, drive days, failures and AFR, but no interval column [1]. You can estimate one yourself with the exact Poisson interval, which is a standard formula [13]. It assumes failures are independent. Backblaze reports that a model's failures can come from separate mechanical problems that hit only some drives [3], so real uncertainty is larger than the formula says. Treat the result as a lower bound on doubt.
8. A data center is not your living room. Backblaze says it plainly: the economics of drives and the measures to keep them spinning in a data center are fundamentally different from a consumer case. Its drives are always on and run until they give up. It also notes that most people have no alerts for drive health at home [7]. Backblaze also writes that its reported failure rates are a function of people actively managing failure and risk, and that its software layer largely prevents data loss [3]. Those are not your conditions. A university study of disk replacements saw little difference between SCSI, FC and SATA drives and concluded that operating conditions may matter more than the drive type [10]. One more gap: the Q2 2026 report says no drive in its tables is SMR (shingled recording) [1]. If you shop for an SMR disk, the table has no data on it, so check the maker's spec sheet for CMR or SMR.
9. Models in the dataset are not the models on the shelf. Backblaze buys in large lots, for example 1,200 drives that filled one storage vault [6]. Some models in the tables are old and no longer sold. Part numbers also change: in 2023 Backblaze wrote that its HGST 12 TB models can still be found but were relabeled as Western Digital with alternate model numbers, and that whether they changed internally is not known to it [5]. Backblaze's own page says that even models from the same maker vary in reliability [2]. We did not check which rows are for sale where you live.
10. What a percentage means for one disk. A model with a 1% AFR means about one drive in 100 fails in a year. In Backblaze's words: "If you have exactly one drive, your personal annualized failure rate is 100%." Its conclusion is to always have a backup and to test it [6]. Datasheets are also optimistic. A large study found datasheet values (MTTF of 1,000,000 to 1,500,000 hours, which means an AFR of at most 0.88%) against field replacement rates that typically exceeded 1%, with 2 to 4% common and up to 13% on some systems [10]. The numbers are good for comparing groups. They are poor at predicting one disk.
11. SMART tells you something, not everything. Backblaze tracks five SMART values: 5 (reallocated sectors), 187 (reported uncorrectable errors), 188 (command timeout), 197 (current pending sectors) and 198 (uncorrectable sectors). In its 2016 data, 76.7% of failed drives had at least one of them above zero shortly before failing, against 4.2% of working drives. That also means 23.3% of failed drives showed no warning [8]. Google's disk study found that over 56% of failed drives had no count in its four strong SMART signals and concluded that SMART alone is unlikely to predict individual failures [9]. A drive can also fail a self-test while its overall SMART status still reads PASSED [11]. Read the single values and run the self-tests regularly.
12. Warranty is the part you control. A warranty is the maker's promise, and it often depends on where you bought. Seagate, for example, tells buyers to claim through the authorized distributor or reseller they bought from, within the same region [15]. In Germany, the legal Gewährleistung against the seller is separate from a maker's voluntary Garantie. For new goods it runs two years [14]. Rules differ in other countries, and this is general information, not legal advice. Read the term for your product line on the maker's page and keep the receipt.

A homelab example

The Bytearium homelab runs rack servers (HPE ProLiant DL360 Gen9 and DL380 Gen10) and a QNAP NAS, as of 14 to 18 Oct 2026. We have not recorded which disks sit in them, how old they are or what their SMART values say, so we say nothing about that. What follows is a general walkthrough for your own Linux machine, not a report from ours.

First, look up your model in the raw data. Download the quarterly ZIP from Backblaze's Drive Stats page [2], unpack it, and check the column order, because schemas change. The schema file lists date, serial_number, model, capacity_bytes and failure as the first five columns (its header says "Last Updated Q2 2024") [2]:

head -1 2026-06-30.csv
awk -F, -v m="ST12000NM0008" 'FNR>1 && $3==m {days++; fails+=$5} END {printf "%d drive days, %d failures, AFR %.2f%%\n", days, fails, fails/days*365*100}' 2026-0[4-6]-*.csv

Replace the model with the one printed on your drive's label. If the awk line returns nothing, the model is not in the dataset for that quarter. Second, add the doubt Backblaze's table lacks. The exact Poisson interval uses chi-squared quantiles [13]. Here it is for 22 failures in 86,092 drive days (this snippet needs scipy):

python3 -c "from scipy.stats import chi2; f=22; d=86092; print(chi2.ppf(.025,2*f)/2/(d/365)*100, chi2.ppf(.975,2*(f+1))/2/(d/365)*100)"

That prints about 5.85 and 14.12. The table shows what the method gives for four Q2 2026 rows [1]. The ranges are our calculation. They are rough because failures cluster. The rows are examples for the arithmetic, not a ranking.

Read it like this: 3.17% and 4.07% cannot be told apart, because their ranges overlap. A zero is not a promise. With 21,499 drive days, zero failures still allows a true rate above 6% (3.6889 is the 95% upper limit for a count of zero [13], divided by 58.9 drive-years).

Third, check the disk itself. This is for ATA/SATA drives. Behind a RAID controller you may need the -d option [12], and SAS and NVMe drives print different output:

sudo smartctl -H /dev/sda
sudo smartctl -A /dev/sda | grep -E '^ *(5|187|188|197|198) '
sudo smartctl -t long /dev/sda
sudo smartctl -l selftest /dev/sda

-H prints the overall verdict, -A the attribute table, -t long starts the extended self-test (tens of minutes to several hours), and -l selftest shows the result [12]. The smartmontools FAQ recommends regular long tests, for example weekly, and smartctl -x -a for a full report [11]. Write down the raw values on day one. A rising number matters more than a single reading. Do not forget the part that holds even when every number looks fine: a second and third copy of your data, as described in our guides on the 3-2-1 rule and on restic and Borg.

Checklist

• Find the exact model number on the drive's label and look it up in the newest Backblaze table. If it is missing, write down that the data say nothing about it.
• Read drive count, average age and drive days before you read the AFR.
• Work out what one failure does to the AFR: 365 divided by drive days, times 100.
• Compare the quarterly row with the lifetime row of the same model, and check the ages in both.
• Calculate a rough 95% range for any row below about 500,000 drive days in a quarter, the bar Backblaze named in 2022 [4], and treat overlapping ranges as a tie.
• Check the maker's spec sheet for CMR or SMR, since the Backblaze tables contain no SMR drive [1].
• Check the warranty term of your product line, the authorized sellers and your statutory rights, and keep the receipt.
• Run smartctl -H, -A and a long self-test when the disk arrives, and record SMART 5, 187, 188, 197 and 198.
• Schedule long self-tests and look at the attributes again, for example weekly [11].
• Spread the drives of one array over more than one purchase when you can. This is our assessment: studies see correlated replacements and have found bad batches [10].
• Keep three copies, two media, one offsite, and restore a test file on a second machine before you need it.