RAID · NAS · Servers
On a broken array, the most dangerous button is Rebuild.
A rebuild writes to every disk in the set. Run with the wrong member order, a stale disk, or a misreading of which drive failed, it overwrites the data it was meant to restore. Everything on this page follows from one rule: nothing writes to the original disks.
Why rebuilds destroy data
A healthy rebuild is routine: one disk is replaced and the controller recalculates its contents from the others. The versions that reach us are the other kind. A rebuild started with members in the wrong order writes parity calculated from a wrong layout across every disk. A rebuild onto the wrong disk erases a good member. A rebuild that includes a stale disk, one that dropped out of the array weeks earlier, folds an out-of-date history into the live data. And a rebuild’s sustained full-surface read is the heaviest load an ageing disk ever sees, which is why the second drive of a tired set so often fails partway through one.
None of that is operator carelessness. Controllers offer the rebuild as the obvious next step, mislabel members, and lose their configuration in the same event that broke the array. The protection is procedural: from the moment the data matters more than the uptime, nothing gets to write to those disks.
Reconstruction, from images
Each member disk is imaged individually, read-only, with error-aware tooling that works around bad regions rather than hammering them. From then on the originals sit on a shelf, and every attempt, every wrong guess and every success happens on the images, where a wrong guess costs nothing.
The array’s real geometry is then determined from the data itself: member order, stripe size, rotation, parity placement and delay, and which member went stale first, established from sequence numbers and from the contents rather than from a controller’s opinion. The array is assembled virtually from the images, the filesystem opened on top, and the files verified before anyone talks about success.
Any layout, any filesystem
RAID 0, 1, 5, 6 and 10, mdadm and LVM layers, and the vendor arrangements built on them, Synology’s SHR among them. On top of those, the filesystem: ext4, Btrfs, ZFS, NTFS, APFS, and VMFS with the virtual machine disks inside it, each of which is a nested recovery of its own. An iSCSI LUN on a NAS is a filesystem inside a file inside a filesystem, and it gets unwrapped in that order.
The stale member problem
Arrays run degraded for months without anyone noticing, because the alert went to an inbox nobody reads. When a second disk falters, the set that arrives contains one member frozen weeks in the past. Include it as current and the volume assembles, mounts, and quietly serves a mixture of two points in time, corruption that looks like success. Part of the geometry work is dating every member and excluding the stale one from any region where a fresher source exists.
When a member is physically dead
Parity exists for exactly this. In a RAID 5 with one unreadable member, the missing disk’s contents are computed from the survivors, and RAID 6 tolerates two. The dead disk still gets assessed free, because a member that can be even partially imaged narrows the gaps, and if its failure is mechanical, heads or motor, that disk is referred out with our findings in writing while reconstruction continues from the members that read. A member that clicks gets powered off, not retried.
Where it ends
More dead members than the parity covers, none of them imageable, is the end: a RAID 5 short two members, or a RAID 6 short three, has holes no mathematics fills. RAID 0 keeps no second copy of anything, so a member that cannot be imaged takes its stripes with it, leaving every large file holed at regular intervals. A completed rebuild or re-initialisation that wrote across the full set may have replaced data outright; regions outside what it wrote can survive, and the assessment maps which. And an encrypted volume without its key or passphrase remains exactly that, here and everywhere else. Bring whatever key material exists, including the recovery key file the NAS offered to download when the encrypted share was created.
For installers, MSPs and IT managers
Most arrays reach us through the client’s IT provider or installer, and the arrangement stays that way: you keep the client, we report to you, findings in writing at every step. If a client’s array is degraded right now and the option on the table is a rebuild, the free step is a phone call before anyone presses it.
Common questions
The NAS says the volume has crashed. Are the files gone?
Usually not. Crashed means the NAS can no longer assemble the volume to its own satisfaction, and the causes, a dropped member, lost configuration, a filesystem fault, tend to leave the data itself intact on the disks. The risk from that moment is what gets tried next on the originals. Imaged and reconstructed off the original disks, these cases end well more often than the error message suggests.
A rebuild has already run. Is there anything left?
Sometimes, and the assessment can say which parts. A rebuild damages what it writes, and it does not write everywhere: an aborted rebuild has a frontier, and data beyond it may be untouched. A completed wrong-order rebuild is the worst case. We map what was written from the images and tell you what survives before any quote.
Two disks have failed in a RAID 5. Is that the end?
It depends on whether one of them can still be imaged. Parity covers one missing member, so the set is recoverable if the second failed disk reads even partially, because whatever it gives up narrows the gaps. Two members that cannot be read at all leave holes no calculation fills, and you hear that at the free assessment rather than after weeks of billing.
Do you need the whole NAS, or the disks? And what does it cost?
The disks carry everything, so the enclosure is rarely needed, though send it if that is easy, because its configuration can shorten the geometry work. Label the disks by bay before removal. RAID and NAS work starts at $690, quoted fixed after a free assessment, and no recovery, no fee applies to the data agreed in writing at quote time.