RAID 5 · one disk of parity
We see this one most weeks. A disk drops out, nobody sees the alert, a spare goes in, and the rebuild halts when the next member will not read right through. Sets reach us from edit suites near Cardiff Bay, professional offices in the city and workshops up the Valleys. It is usually less bad than it looks: failed is a verdict from the card, not a post-mortem, and whatever we send back is read off a clone, not off the disk in your box.
Describe the symptoms to an engineer
0800 6890668
| What shows up | What has gone wrong | Do this now |
|---|---|---|
| The controller lists a member as missing | The parity is used up; one more fault and the volume goes | Power it off before the next one |
| A rebuild ran for hours, then halted | Reading every sector has turned up the soft spot on another disk of the same vintage — how it normally ends | Leave it powered down. No second attempt |
| Two members are flagged failed | Almost always one dead disk and one holding a handful of unreadable sectors, not a pair of dead disks | Send all of them; the clones sort out which is which |
| A dropped member was forced back into the set | Stale blocks are now sitting among live stripes | Go no further |
| A new card will not recognise the array | The argument is over metadata — the files are still where they were | A routine reassembly |
| Recovery software has already been run over the members | Something has written to the original disks | Send them in, and tell us which tool it was |
The diagnosis costs you nothing. Every item that arrives is logged under a case number of its own, and an engineer works out what has actually failed — then tells you plainly which files can come back and which cannot. The price follows that: one fixed figure, in writing, and no work begins until you have read it and said go ahead.
Each disk is copied in full onto our own storage, the ones the card condemned included. Weak areas are logged as the copy goes and the most stubborn are saved for a final pass. Nothing is written back to your disks at any stage.
Disk order, stripe size and the way parity rotates come from two places: the metadata the card left on each member, and the pattern of the data itself. None of it is guesswork, and none of it needs your controller.
The set is put back together over the images. Every block is taken from whichever copy read it cleanest, and where no copy could manage it, parity fills the gap — which is how files stranded on a written-off member come back. We then repair the file system and mount it to check.
Nothing is billed while you are still deciding. A complete listing of what came off the drive reaches you first, and the invoice only follows your yes. Files travel back on media bought new for the job, return postage ours, and we do not close it here until every one of them opens on the machine you will use.
Work it out once, then leave it alone: a consumer disk is published at one URE — an unrecoverable read error — for each 1014 bits it reads, roughly one bad sector per 12.5TB pulled off it. Enterprise disks are specified at 1015. Rebuilding a four-bay array of 16TB drives means asking three survivors for around 48TB without a stumble, so this is not a paper worry. Plenty of disks also run well past their rating. Both are true, and both lead to the same move: clone first, and a rebuild costs you nothing.
Back in June, disk two left the set without anyone noticing. Then, in August, disk four gave out during a rebuild, and from the outside it looked as though a pair had failed together. They had not. On the imager the fourth showed nineteen bad sectors and no other damage, so three sound members plus that near-sound fourth were enough to put the stripe back. Year-end filings went in on time.
Rarely. A controller drops a disk on its own criteria, and often that means a handful of sectors that answered too slowly rather than a drive that has died. With every member imaged onto our storage, whatever one clone cannot give up is reconstructed from the parity sitting on the rest. Two genuinely dead disks are unusual. Two disks a card has ejected are routine.
Not until the disks have been imaged. A rebuild asks every surviving member for a full end-to-end read, which is precisely what finishes off a disk that is already tired. It also writes over the parity we would use to put the volume back. Once the clones exist, you can rebuild as often as you like.
A URE is a sector the drive has given up on once its own error correction has run out. Consumer disks are published at roughly one per 12.5TB read, enterprise disks at one per 125TB. Treat that as the floor the manufacturer will defend, not a clock ticking down. Modern firmware handles a URE by puncturing that single stripe and letting the rebuild carry on.
No, the disks are enough. Each member holds the card's own map of the array, and where that map is damaged the layout can be worked back out of the data. We assemble the set in software over the clones. Nothing is ever written to the disks you send.
A failing drive has only so many starts left in it. Spend none of them. Keep the power off, and the free diagnosis will tell you what is still readable.