Three grades, one word each
Manufacturer recertified means the drive went back to WD, Seagate or Toshiba — an RMA return, an overstock unit, or a hyperscale buyback — was tested on the factory’s own equipment, had its SMART counters reset, and got a new label (WD’s are white-label “Recertified”; Seagate’s say “Factory Recertified” and often ship green-bordered). This is the only grade where anyone with the drive’s design documentation vouched for it.
Seller refurbished means whatever the seller wants it to mean. At the good end, a serious reseller ran a full-surface write/read pass and rejects anything with reallocated sectors. At the bad end, it means “it spun up when we plugged it in.” The label carries no information; the seller’s return policy and feedback history carry all of it.
Used pulls are drives decommissioned from data centres and sold as-is, usually with SMART data intact. Counterintuitively, that honesty is a feature: a pull showing 30,000 power-on hours and zero reallocated sectors tells you more than a “refurbished” drive showing zero of everything. Enterprise drives are rated for five years of 24/7 duty; a three-year pull from a climate-controlled hall has led an easier life than most desktop drives.
The warranty you actually get
The original factory warranty follows the first sale and is dead on all three grades — a serial-number check on the vendor’s site will say “out of warranty” or “product of recertification.” What replaces it: manufacturer-recertified stock typically carries six months to two years, honoured by the vendor or its authorised reseller (Seagate’s factory-recert drives through authorised channels usually carry one to two years). Seller-refurbished and pulls carry only the seller’s own promise — commonly 30–90 days, sometimes a “5-year warranty” that is really a store credit policy from a company that may not exist in five years. Before paying a premium for a long seller warranty, price it as if it were 90 days and see if the deal still works.
What failure rates to expect
Drive mortality is a bathtub curve: elevated failures in the first months (infant mortality), a long flat floor, then a rise as mechanics wear out. Healthy enterprise drives sit on the floor at roughly 1–2% annualised failure rate — Backblaze publishes fleet numbers every quarter if you want model-level detail. Recertified and pulled drives have already survived infancy, which genuinely works in their favour; against that, RMA-return recerts include an unknown fraction of drives that already misbehaved once, and high-hour pulls are further along the curve. A fair planning number for decent recertified stock is roughly double the new-drive AFR — call it 2–4% — with a wider spread between good and bad batches. Nothing about that is scary in a redundant array with backups. All of it is scary as a single drive holding the only copy.
When the discount is worth it
The break-even is arithmetic, not vibes: a used drive is worth it when the discount exceeds the cost of the extra expected failures and the shorter warranty over your holding period. As a rule of thumb, recertified starts to make sense below about 70% of the new price and gets compelling below 50% — which, for large-capacity SAS, is where the market usually sits. We built the used-vs-new calculator to do this properly: it takes the failure-rate multiplier and warranty gap and tells you the exact used price above which the discount stops being real, using live prices from the SAS and SATA rankings.
What to check on arrival
Test every drive inside the return window, before it holds data. Pull SMART first — smartctl -a /dev/sdX (add -d sat behind some USB bridges). The fields that matter:
- Power_On_Hours (attribute 9) — does it match the story you were sold? 8,760 hours is one year of 24/7. Exactly zero on a “refurbished” drive means the counters were reset, not that the drive is new.
- Reallocated_Sector_Ct (5), Current_Pending_Sector (197) and Offline_Uncorrectable (198) — the media-health trio. Any nonzero value on arrival is a return, full stop. These are the strongest SMART predictors of imminent failure.
- Command_Timeout (188) and UDMA_CRC_Error_Count (199) — nonzero CRC errors usually mean a bad cable or backplane rather than a bad drive, but note the baseline so you can tell if it climbs.
- On helium drives, attribute 22 (helium level) should read at or near full; on SAS drives, read the grown defect list and error counters instead (smartctl -x or sg_logs) — a nonzero grown-defect count is the SAS equivalent of reallocated sectors.
Then make the drive prove itself: a SMART long self-test (smartctl -t long) at minimum, a full write-and-verify pass (badblocks -wsv, on an empty drive only) if the return window allows the day or two it takes at 20 TB. Re-read SMART afterwards — a drive that grows pending sectors during one surface pass was going to fail on your data instead. Do this every time, on every grade, including new drives. The label tells you who tested it last; the test tells you the truth.