RAID

Start with the layout, not the number of drives. RAID 0 stripes data without protecting it; RAID 1 keeps multiple complete copies; RAID 5 and RAID 6 use parity; conventional RAID 10 stripes over mirror groups. These choices trade capacity for different recovery paths. [1] [2] [3] [4] [5]

For N equal-sized active members, each contributing capacity S, the table gives ideal data capacity. Spares, array metadata and filesystem overhead are excluded. RAID 1 means one N-way mirror here; RAID 10 means fixed two-member mirror pairs with an even N of at least four. [6] [5]

Four 8 TB members illustrate the distinction: RAID 0 provides 32 TB, a four-way RAID 1 mirror 8 TB, RAID 5 24 TB, RAID 6 16 TB and paired RAID 10 16 TB before overhead. These follow from the table's formulas; the RAID 1 row is not two separate mirrors.

Equal member sizes are an assumption, not a universal RAID 0 capacity rule. Linux MD can continue into narrower stripe zones after smaller members run out of space, using remaining capacity on larger members. A full RAID 1 mirror instead needs room for the complete logical data on every participating member. Verify the implementation and usable member sizes before applying a formula. [10]

Failure position matters in paired RAID 10. With pairs A/B and C/D, losing A and C leaves one copy in each pair; losing A and B removes both copies of that pair's data. Counting failed drives alone cannot distinguish these cases. Linux also supports near, far and offset RAID10 layouts, including odd member counts, so the fixed-pair example must not be applied to every RAID10 implementation. [5] [7]

Parity capacity is an overhead equivalent, not necessarily a dedicated drive. RAID 5 rotates parity across members; RAID 6 layout details vary. Do not infer where parity lives merely from subtracting one or two members in a capacity formula. [3] [7]

A stripe unit, also called a strip or stripe element, is the contiguous portion on one member. It is not the full stripe. For four active members with 512 KiB units, full-stripe user data is 2 MiB in RAID 0, 1.5 MiB in RAID 5 and 1 MiB in conventional RAID 6. The parity layouts still occupy 2 MiB across the members; their user-data totals exclude parity. [9] [10] [7]

Request length alone does not establish stripe coverage. A stripe-aligned 1540 KiB RAID 0 request touches all four members: three 512 KiB chunks plus 4 KiB of the fourth. It does not fill the 2 MiB stripe. In the four-member RAID 5 example, a 1.5 MiB request starting 4 KiB past a data-stripe boundary spans two partial stripes. These are address-mapping calculations; the parity layer may aggregate separate writes before issuing a full-stripe write. [8] [14]

Mirroring does not promise a fixed read-speed multiplier. Linux MD documents read balancing across members but distinguishes a single sequential stream from concurrent streams or random reads. Writes must update the active copies; workload, layout and implementation determine the performance result. [10]

A configured mirror is not necessarily a synchronized mirror. By default, mdadm starts an initial resync when creating a RAID 1 array. With optional write-behind enabled, Linux MD can report write completion before designated write-mostly members have caught up. A surviving device must contain the needed current data; counting installed members alone does not establish that condition. [11] [10]

Consistency is not integrity. Krioukov and colleagues showed how recomputing parity from unvalidated data can make corrupted data and parity agree. Their FAST 2008 model explores one stripe with one injected error and collision-free protection primitives; it is not a present-day fleet failure-rate measurement or a reason to disable scrubbing. Validation must cover maintenance reads as well as application reads. [12]

Interpret scrub counters at their actual granularity. Linux MD mismatch_cnt counts sectors rewritten, or that would be rewritten during check, in units that can exceed the damaged region. The mdadm manual gives an example where one failed 64 KiB comparison adds 128 sectors. This is not a count of damaged files or independent failures, and check is not guaranteed to be read-only for every RAID level. [13] [10]

Redundancy and backup address different failures. A healthy mirror can faithfully preserve an unwanted overwrite. Recovery needs suitable retained copies, isolation and a tested restoration path; RAID alone does not provide those properties. [15]

The naming has a history: the December 1987 Berkeley report and the 1988 SIGMOD paper are titled A Case for Redundant Arrays of Inexpensive Disks (RAID). Their publication dates identify distinct records, not the first invention of every mechanism now grouped under RAID. [16] [17]