
In short
- Read before you write: mdadm --examine on every member, and note the result.
- --assemble --force is a calculated risk; --create is a last resort that can destroy data.
- Work on clones or copy-on-write overlays, never on the only copy.
Rule number one: copy before you experiment
If the data matters, image every member first (with GNU ddrescue on a healthy machine, or in a lab if a disk is failing). If you cannot clone, at least work through copy-on-write overlays: device-mapper snapshots that send every write to a separate file, so the original disks are never modified. Every experiment below should happen on clones or overlays.
Read the superblocks
Each member of an md array carries a superblock describing the array. Start with:
cat /proc/mdstatto see what the kernel has assembled, if anything.mdadm --examine /dev/sdX1on every member partition.
Note for each member:
| Field | Why it matters |
|---|---|
| Array UUID | All members of one array share it |
| Raid Level, Raid Devices | Confirms the layout |
| Chunk Size, Layout | Stripe size and parity rotation |
| Device Role | The position of this disk in the array |
| Events | A counter increased on each array change: the most important clue |
| Update Time | When this member was last written |
| Array State | How this member saw the array the last time it was updated |
Reading the event counter
Members with the same (or nearly the same) event count are in sync. A member far behind dropped out earlier and holds stale data. In a RAID 5 with four disks where one member is thousands of events behind and another is a few events behind, the best assembly usually uses the three most recent members and leaves the stale one out.
Assembling read-only
Try first: mdadm --assemble --readonly /dev/md0 followed by the members you have chosen. If mdadm refuses because the event counts differ slightly, --force tells it to accept the most recent members and update the lagging superblock. That is a write: do it on overlays, and only when the difference is small and you understand why.
Then mount the file system read-only (for ext4, mount -o ro; for XFS add norecovery so the log is not replayed) and copy the data off before doing anything else.
Why --create is dangerous
mdadm --create with --assume-clean writes new superblocks with the parameters you give. If the chunk size, layout, member order, metadata version or data offset differs from the original even slightly, the array assembles into garbage, and the original superblocks are gone. Newer mdadm versions also use a different default data offset than older ones, a classic trap on arrays created years ago. It is a last resort for specialists, on overlays, with the old --examine output in front of you.
When to stop
- A member makes noise, disappears from the bus or shows rapidly growing reallocated or pending sectors.
- Two members have very different event counts and you are not sure which one is current.
- The file system will not mount read-only.
At that point, every extra attempt reduces what can be recovered. Our RAID data recovery service rebuilds the array from clones, in software, without touching your disks. See also RAID 5 rebuild failed: what happened and what to do now.
Frequently asked questions
Is mdadm --assemble --force safe?
It is relatively safe when the event counts differ by a few units and the disks are healthy, but it writes to the superblocks. Use overlays or clones whenever the data matters.
What does a big difference in the Events field mean?
That member left the array long before the others. Its data is out of date and should normally be left out of the assembly.
Is your data important?
Do not try anything that could make the fault worse. The diagnosis is free and you get a fixed quote before any work starts.



