RAID 5 · one parity drive
Parity in RAID 5 covers one lost disk and nothing beyond it. What finishes the second disk, more often than not, is the rebuild itself — because a rebuild demands a complete uninterrupted read from drives bought in one carton and run at one temperature ever since. Sets reach the Reading bench from distribution units at Theale, small software teams at Earley, and the four-bay box in the stores at a Woodley trade counter. The front panel nearly always reads worse than the truth, because failed is a verdict a controller reaches, not a diagnosis. From £500 + VAT, one figure in writing after a free look, and the disks you send go home exactly as they came.
An engineer answers, and looking is free
0800 6890668
| What you are seeing | What is likely wrong | What happens next |
|---|---|---|
| The controller reports a member missing | Your one disk of parity is now spent, and a second fault takes the volume with it | Shut it down before another member fails |
| A rebuild ran half the night, then halted | A full read of the others met a soft patch on a disk of the same age — the usual ending | Keep it switched off. Do not retry |
| Two members are showing as failed | Almost always a finished disk sitting next to one that is merely slow over a handful of sectors — not two of a kind | Post them all; the images separate them |
| A rejected disk was forced back into the set | Out-of-date blocks now sit among live stripes | Touch nothing else |
| A replacement controller cannot see the set | This is a quarrel over metadata, and not one of your files has moved | A normal reassembly |
| Recovery software has been run at the members | Something has been written to your original disks | Send them anyway, and name the tool that ran |
A case number goes on your device the day it reaches us, and an engineer then finds the real fault — free, and before anything else happens. You are told plainly what will come back and what will not, with one fixed figure in writing beside it. Nothing chargeable starts until you have read that and said yes.
Every member is copied in full onto our storage, the ones the card wrote off included. Slow reads are logged as they happen and the worst regions are left until last. Not one byte goes back to the disks you posted.
Three things have to be settled: the bay each member came out of, the size of a stripe, and the direction parity travels round the set. All three come out of what each disk recorded about itself, checked against the pattern of your own data. Nothing here is estimated.
The array is then assembled on top of the images, never on the disks. Every block comes from whichever copy delivered it most cleanly; where no copy did, parity fills the gap — which is why data living on a member the controller had written off still arrives home. After that the file system is repaired, and the rebuilt volume mounted and checked before anything is released.
You see the file list first, then decide. Nothing is invoiced until you have said yes, and on most jobs there is no fee at all if the data does not come back. What we pull off goes onto media bought in for your job, and the postage home is ours. The case stays open on the bench until you tell us the files open on your own machine.
The arithmetic, done once: a consumer disk is rated at one unrecoverable read error in 1014 bits, which is one awkward sector every 12.5TB or so; enterprise disks are ten times better, at 1015. Put four 16TB drives in a chassis and a rebuild wants roughly 48TB of faultless reading from the three that remain, so this is arithmetic rather than scaremongering. Set against it, disks in the field routinely run well past what they are rated for. Both are true, and both point the same way: images first, rebuild afterwards, because once the copies exist a rebuild costs you nothing.
It read as two disks failing at once. It was not. One member had quietly left the set months before, unnoticed, and a second stopped midway through the rebuild that followed. Imaging put only nineteen bad sectors on that second disk, so three sound members and one very nearly sound were enough between them to work the stripe out again. The year-end filing went in on time.
Usually not. Controllers drop members for reasons of their own, and a handful of sectors answering a fraction too slowly will do it. A disk dropped from an array is not the same thing as a disk that is finished. Each member gets imaged to our storage before anything else; whatever a single image withholds is then reconstructed from the parity the others still carry. A controller marking two members bad makes for an ordinary week's work; two members genuinely beyond imaging is uncommon.
Not before the copies exist. A rebuild instructs every remaining member to read itself from end to end — exactly the demand a tired disk cannot meet — and it lays new parity across the old parity that would have put your volume back together. Image the set first and you may then rebuild as often as you like, since nothing you do afterwards can alter the copies.
It is a sector that a drive's internal error correction cannot put right, at which point the drive halts and says so rather than hand back something wrong. Consumer disks are specified at roughly one such event per 12.5TB moved; enterprise disks at one per 125TB. Treat those as guaranteed minimums a manufacturer will defend, not as a countdown to a date in your diary. Modern array firmware meets a URE by punching out the stripe and continuing.
No — the disks alone are enough. Each member holds its own account of how the set was built, and where that account is damaged, the shape of your own data says the same thing. The set is then assembled in software above the images, nothing is written to what you posted, and your controller stays in the machine.
You pay nothing for the look. What comes back is a list — the files that opened, the ones that did not, and one fixed figure to finish the job. Leave it switched off until you have that in writing.