Hard drive failure statistics dataset: how storage hardware actually dies, delivered as rows

Exactly one named record measures what the phrase promises. Backblaze Hard Drive Test Data (Drive Stats Raw Dataset) snapshots every operational drive in Backblaze's production fleet once a day - since April 2013 without a break - and the archive now stands at roughly 341,000 active drives and about 530 million cumulative drive-days. Within the systems-software slice of the Datadory catalog (25 pooled records, 24 primary), it is the only record whose unit of observation is the individual device-day, and it scores 10 on our quality rubric - one of eight perfect scores in the slice.

Everything else nearby measures a different quantity. Valve's Steam Hardware & Software Survey reports how much total and free disk space surveyed machines carry, not how often those drives die. The Observatory of Economic Complexity counts dollars rather than failures: $60.2 billion of HS 847170 storage-device trade in 2024. Icecat documents specifications per GTIN. If the question is how often a given drive model fails in service, the drive-day panel is the only direct measurement available - no vendor datasheet publishes an observed fleet history at this grain.

Three zones per row - identity, placement, telemetry - flat and identical in shape whether the day falls in 2013 or last month:

date           : 2026-02-15
serial_number  : 000a43e7dee60010
model          : TOSHIBA MG07ACA14TA
capacity_bytes : 14000519643136
failure        : 0
datacenter     : ams5
cluster_id     : 031
vault_id       : 2017
pod_id         : 05
smart_5_raw    : 0
smart_9_raw    : 46458

Eleven fixed columns carry identity, placement and the label; a wildcard pair carries telemetry - one column per SMART attribute ID the drive reports, in raw and vendor-normalized variants. Because different models expose different attribute sets, the telemetry zone breathes with the fleet. Two figures show the telemetry zone doing its job: smart_5_raw (reallocated sector count) parked at zero while smart_9_raw (power-on hours) reads 46,458 on one drive and 2,798 on another of the same model family - two units at very different ages contributing identical-shaped rows.

The failure flag is why the panel compounds. It reads 1 only on the day a drive left service, which makes every other day an exposure observation - the denominator side of any failure-rate calculation. A boolean marks pre-2018 legacy-layout records, so early and current vintages sit in one panel without guessing which schema you are reading.

The arithmetic is two aggregates, and the panel carries both sides natively:

1. Count exposures. Drive-days per model equal surviving rows per model per day - the denominator.
2. Count exits. Failures equal rows where failure = 1 - the numerator.
3. Annualize. Failures over drive-days, scaled to a year, yields the annualized failure rate for whatever cohort you drew.
4. Slice. Repeat per model, capacity point or vintage. Lifetime across the full ledger lands near 1.39%; a recent quarter contributes roughly 30 million drive-days holding about 1,030 failures.
5. Model beyond the summary. Survival curves, time-to-failure distributions and next-day classifiers all hang on the same two columns plus SMART trajectories as time-varying covariates - the per-day rows support far more than any published summary table can.

The step teams skip is the one that breaks the result: counting only failures while ignoring the surviving days inflates every rate. Exposure-time discipline down to the individual drive-day is what separates an annualized failure rate from an anecdote with a decimal point.

Four other cataloged records orbit the failure question without answering it directly. They differ by what they measure, so treat them as complements rather than substitutes:

Four constraints shape any estimate built on this corpus.

Fleet composition. This is an operational-fleet sample skewed toward high-capacity enterprise and NAS drives, not a consumer retail sample, so absolute failure rates do not transfer to laptop workloads. The bias cuts the right way for procurement work - the fleet resembles what a serious storage buyer deploys, not a random drawer of leftover disks.

Placement, not geography. Rows locate drives in specific facilities - US sites plus international ones such as Amsterdam (ams5) - which makes this a reliability panel rather than a geographic survey.

Exit-only failure signal. The flag marks the day a drive left service; it is a termination event, not a graded symptom history, so degradation studies lean on SMART counter trajectories instead.

Vintage drift and thin event counts. Early vintages are effectively HDD-only, and SSD representation becomes statistically meaningful only in later years, so solid-state claims need late-vintage windows. SMART column sets shift as models rotate through the fleet, with two documented layouts spanning the timeline. And the signal base is finite: sub-model conclusions get noisy quickly below the highest-volume SKUs, because roughly 1,030 terminal events per quarter spread thin once sliced past the top SKUs.

Failure statistics earn more when crossed with installed-base and supply data, and the surrounding slice supplies four natural joins.

Installed base: the Steam Hardware & Software Survey prints monthly distributions for OS version, CPU cores, VRAM and total versus free disk space - the July 2026 edition showed Windows 11 64 bit at 70.26%, Linux at 4.01% and macOS at 2.32% - which tells you whose drives your failure curves should describe.

Support windows: the endoflife.date Product Lifecycle Catalog tracks end-of-life dates for 464 products, typically split into 3-30 release cycles, so replacement planning can line drive age against OS support cutoffs - failure curves tell you when drives die in practice, lifecycle dates tell you when vendor support disappears in principle.

Supply side: the OEC Hard Disk Drives Trade Profile sizes the market - $60.2 billion of HS 847170 trade in 2024, up from $51.3 billion in 2023, led by Thai exports of $16.5 billion.

Parts catalog: Icecat Open Catalog carries 30,266,450 datasheets across 29,756 brands with GTIN mapping, so model strings in the telemetry resolve onto sellable SKUs.

• Reliability engineering and survival analysis. A failure flag plus per-drive exposure turns half a billion drive-days into hazard curves and time-to-failure distributions at per-model resolution; rare-event statistics need exactly this much runway.
• Predictive-maintenance ML. SMART counters as features, the next-day flag as a label - a supervised problem with roughly half a billion rows carrying explicit outcomes, unusually large for hardware. The class imbalance is brutal and honest, which is precisely what makes trained models transfer.
• Procurement benchmarking. Buyers compare annualized failure rate by model, capacity and vintage before committing purchase-order volume, replacing anecdotes and vendor datasheets with observed fleet behavior.
• Capacity and lifecycle planning. Watching which models and capacities exit service, and when, informs spares pools, warranty windows and replacement budgets years out.

Within Datadory's catalog these workflows rank among the strongest fits in systems software: the slice's primaries average about 8.4 on the quality rubric against a catalog-wide mean of 7.81, and the drive-day panel is the record data-science teams cite first (data scientists view).

Files, feeds, or your warehouse. Daily, weekly, or hourly - your call.

Cohort hygiene comes standard. Snapshots pin by date, because a report re-run next month should reproduce, and deliveries land keyed on serial number and date so each pull joins cleanly to the last - and to whatever else you already hold. Derived folds ship on request: model-level annualized failure rollups, placement hierarchies resolved into clean dimension tables, wide or long reshapes, scoped extracts filtered to the models, capacities, facilities and year ranges you name.

Start with the systems software data guide for how drive telemetry fits the wider registry, security-feed and benchmark landscape, and the systems-software data hub for field dictionaries and sample rows across all 25 pooled records. For adjacent pipelines, the open source vulnerability feed covers the advisory layer and the software end-of-life tracker pairs drive aging with support-window data. The record-level detail lives on the Backblaze Hard Drive Test Data dataset page; the best systems-software datasets ranking places it within the slice's 25 pooled records. When you want rows rather than reading, request a sample and the failure panel arrives decoded, aligned and pinned to the dates you name.