mdadm RAID concepts - levels, layouts, and failure modes · Linux

RAID (Redundant Array of Independent Disks) provides redundancy and/or performance by combining multiple physical disks into a logical unit. The choice of RAID level depends on the workload and the failure-tolerance budget.

The RAID levels

Level	Min disks	Fault tolerance	Capacity	Write hole	Use for
RAID 0	2	0 disks	N * size	No parity, no redundancy either	Temp scratch, never production data
RAID 1	2	(N-1) disks	1 disk	None	Boot disks, small critical data
RAID 5	3	1 disk	(N-1) * size	Yes - needs PPL, a write journal or a healthy BBU	General-purpose data, only once the write hole is closed
RAID 6	4	2 disks	(N-2) * size	Yes - narrower window, not closed	Large data, longer rebuilds
RAID 10	4	(N/2) disks	N/2 * size	None	High-performance, high-reliability, anything that must not corrupt silently

The write hole column is the one most selection tables omit, and omitting it is how RAID5 gets chosen for a database. Updating a parity stripe is two non-atomic writes - the data chunk and the recomputed parity chunk, on different disks. Lose power between them and parity no longer matches data, yet the array comes back clean. The damage surfaces months later, when a disk fails and md reconstructs from the stale parity and returns plausible, wrong data with no error anywhere.

Mirroring has no write hole because there is no parity to recompute. On a parity level, close the hole explicitly with PPL (--consistency-policy=ppl), a write journal (--write-journal), or a controller with a verified-good BBU - and scrub monthly. The lesson linux-raid5-raid6-parity covers the mechanism and the mitigations in full.

Trade-offs

Read-only / Safemdadm array— md0 is a RAID 10 array composed of four 100 GB disks. Capacity is 200 GB (half the raw total), fault tolerance is up to 2 disks if they are in different mirror groups. The output is from a typical production host.

$ lsblk -o name,size,type,fstype,mdraid

NAME          SIZE TYPE FSTYPE MDRAID
sda          100G disk
sdb          100G disk
sdc          100G disk
sdd          100G disk
md0          200G raid10 ext4

Read-only / Safemdstat— /proc/mdstat shows the array state. [4/4] means all 4 devices are present and operational. [UUUU] shows each device's state: U = up, _ = down. If a device fails, you see [_UUU] for example. The personalities line lists the RAID levels supported by the kernel module.

$ cat /proc/mdstat

Personalities : [raid0] [raid1] [raid10] [raid6] [raid5] [raid4]
md0 : active raid10 sdd[3] sdc[2] sdb[1] sda[0]
204800 blocks super 1.2 512K chunks 2 near-copies [4/4] [UUUU]

Common failure modes

Read-only / Safemdadm detail— mdadm --detail shows the array metadata. State: clean is the normal state. State: degraded means one device is missing. State: recovering means a rebuild is in progress. State: failed means too many devices have failed.

$ mdadm --detail /dev/md0 | head -10

/dev/md0:
Version : 1.2
Creation Time : Mon Jan  1 10:00:00 2026
Raid Level : raid10
Array Size : 204800 (200.00 MiB 200.00 MB)
Used Dev Size : 102400 (100.00 MiB 100.00 MB)
Raid Devices : 4
Total Devices : 4
Persistence : Superblock is persistent
Intent Bitmap : Internal
Update Time : Mon Jan  1 10:00:00 2026
State : clean

When a device fails, the array becomes degraded. The kernel continues to serve reads and writes from the remaining devices. Performance drops because the missing data must be reconstructed. Once you replace the failed disk and add it to the array, mdadm rebuilds the array.

Configuration changeadd new disk to array— mdadm -a adds a new device to the array. The kernel starts a rebuild immediately if the array is degraded. Watch /proc/mdstat for progress: [4/4] [UUUU] means the array is fully rebuilt.

$ mdadm /dev/md0 -a /dev/sde

mdadm: added /dev/sde

The RAID levels

Trade-offs

Common failure modes

Knowledge check