A home NAS builder computes real drive failure odds from Backblaze data

The author, posting on Hacker News as user nosolace, wants their personal data to survive for their entire life, and possibly another generation, so they set out to figure out exactly how reliable a home storage array would have to be to get there. Rather than take manufacturer datasheets at face value, they built a full probabilistic model of data loss risk, redoing an analysis they had done before at a previous job running about 200 TB of networked storage.

The planned architecture: four 8 TB drives, split into two groups of two, each set up as a btrfs RAID1 mirror. One mirror is kept at the author's father's house several states away, with the local and remote copies synchronized automatically. btrfs checksums each roughly 16 KiB data block and can auto repair a corrupted block on read as long as a good copy of it exists somewhere else in the array, a feature zfs also offers. The author also plans to rely on the drives' built in SMART monitoring, and treats two distinct failure modes separately: a whole drive dying, which loses everything the drive held, and a latent sector error, an unreadable spot confined to a roughly 512 byte sector.

Classical reliability engineering describes hardware failure with a bathtub curve, high failure early and late in a device's life, low in the middle. The author argues this shape increasingly does not fit modern hard drives, and cites Backblaze's October 2025 post on annualized failure rates, whose graph is notably not a bathtub. Backblaze reads the changing curve as showing both that drive reliability has improved over roughly the last decade and that the nature of failure itself has shifted; the author calls the resulting year to year variability in failure rates worrying, since it suggests failure is not simply a function of age. They also draw on Bairavasundaram et al.'s 2007 SIGMETRICS analysis of single sector failures and on a study by Schroeder, Damouras and Gill covering drive and sector failure behavior.

Because the existing literature did not give recent failure probabilities for the specific drives the author was buying, they had an AI model called opus 4.7 pull the most recent, 2025, Backblaze dataset and examine drive models similar to the ones being ordered. The author and the model went back and forth on exactly what to look at before settling on tracking new failures at sector granularity. They found that 3 to 5% of drives in the dataset showed some newly appearing condition, though the extracted text cuts off before naming exactly what that condition was. The model noted that the single most common occurrence was an eight sector failure, which it theorized corresponds to one 4 KiB on disk sector, an explanation the author accepted without digging further into the mechanics. The text does not say which company makes opus 4.7.

To eventually estimate expected data loss volumes, the author also sized up their own annual data workload: a 100 GB home directory with nightly backups changing an average of 1 GB each, roughly 100 GB film files watched once or twice a week, and lossless, 24 bit 44.1 kHz FLAC equivalent, music streaming totaling around 50,000 minutes a year per their own listening history. That comes to about 1.5 times ten to the 13th power bits per year alone, or about 3 times ten to the 13th power bits per year once a partner's usage is added in, figures given as printed in the source.

On bitrot and scrubbing, sources generally recommend scrubbing checksummed NAS filesystems regularly so bad sectors are caught and corrected before a resilver becomes necessary. The author notes it is unclear from datasheets alone whether bitrot, data quietly decaying with no read error reported, is a real occurrence on modern magnetic hard disks; they found conflicting reports, and note that Australian government claims about temperature affecting stored data are mostly about magnetic tape rather than hard drives. Without knowing bitrot's rate or mechanism on hard disks, the author calls scrubbing's efficacy unspecified, since it only matters if data can change once written. What regular scrubbing does reliably ensure is that, if a drive is later lost, the surviving drives are known to have had at least one successful read of every sector that drive held.

The author then builds the actual probability model. Data loss over the array's life is treated as the combination of two paths: one drive out of the array failing and a second loss event then occurring among the remaining drives during the resilver window, assumed at two weeks since most of that time goes to procuring a replacement drive rather than the rebuild itself, or all the drives independently developing a latent sector error in the exact same sector. Three modeling assumptions carry the analysis: an array is treated as lost only once, after which it is rebuilt from scratch; drive failure rates are treated as constant with age, since the shape of the failure curve no longer looks strongly age dependent; and failure events on different physical drives are treated as statistically independent. Using a Poisson process, exponential distribution model, the author derives that the probability a single drive survives t years equals e to the negative AFR times t, extended to cover several drives sharing the same annualized failure rate.

The extracted text breaks off mid derivation, in the middle of working out the probability that surviving drives share a latent sector error, before the author reaches any final numeric data loss probability or states a conclusion. No overall result or recommendation is present in the material available for this retelling.

Key facts

• Ordered four 8 TB drives arranged as two btrfs RAID1 mirror pairs, with one mirror kept off site at the author's father's house and synced automatically to the local copy.
• Cites Backblaze's October 2025 annualized failure rate report to argue the classic bathtub curve of hardware failure no longer fits modern hard drives.
• Had an AI model called opus 4.7 pull and examine the 2025 Backblaze dataset, finding 3 to 5% of drives showed a newly appearing condition and that eight sector failures, theorized as a single 4 KiB sector, were the single most common event.
• Built a Poisson process probability model combining whole drive failure, with an assumed two week resilver window, and the risk of multiple drives sharing a latent sector error, but the available text ends before a final data loss probability is reached.
• Estimated personal annual data workload at about 1.5 times ten to the 13th power bits alone, or about 3 times ten to the 13th power bits combined with a partner's usage, from nightly backups, film watching and lossless music streaming.

Why it matters

The post is a rare case of someone actually doing the math on home storage reliability instead of repeating vendor marketing about datasheet annualized failure rates. It treats redundancy, checksums and off site copies as inputs to a real probability model rather than a checklist, and challenges the standard bathtub curve assumption using a fresh industry dataset.

Who it affects

Home lab and self hosted storage enthusiasts, anyone maintaining a personal NAS or planning a long term personal data archive, and readers interested in applying reliability engineering statistics outside a data center context.

How to use it

The concrete recipe described is four drives split into two btrfs RAID1 mirror pairs, one mirror kept at a second physical location and auto synced, checksummed 16 KiB blocks with auto repair on read as long as a good copy exists elsewhere, SMART monitoring, and a periodic, roughly monthly, scrub of the whole array plus a readback immediately after every write.

How solid is it

The analysis rests on a named, recent industry data source, Backblaze's October 2025 failure rate report, plus two peer reviewed academic studies, Bairavasundaram et al. in SIGMETRICS 2007 and a paper by Schroeder, Damouras and Gill, alongside AI assisted examination of the raw dataset. The math itself, a Poisson process model of drive and sector failure, is shown step by step in the post, but the text available for this retelling is truncated mid derivation, before any final data loss probability or conclusion appears.

Risks and caveats

The model's own simplifying assumptions, that drive age can be ignored and that failures across separate physical drives are independent, are stated by the author as reasonable rather than proven. Several concrete details are missing from what could be extracted: the exact Seagate drive model in use, the specific condition behind the '3 to 5% of drives' figure, and any final AFR or data loss probability number, so the piece cannot yet be checked against a stated bottom line.

“i will assume that a given array will only experience data loss once: after the first moment of loss, the array is toast and we have to rebuild it.”

— the author, in the post "P[drive failure]: how reliable is your NAS?"