RAID Levels in Practice: 1, 5, 6 and 10
RAID levels look like a menu but behave like physics. Each is a specific trade between usable capacity, write cost and how much failure you survive — and because drives grew a hundredfold while rebuild speeds did not, advice that was right in 2010 is quietly wrong today. Here is the working engineer's version.
RAID1 / RAID10: pay half, sleep well
Mirrors are the simple, fast answer: every write goes to two drives, reads come from either, and a failure degrades performance barely at all. Rebuild is a straight copy — quick, low-stress, no parity math. RAID10 stripes across mirrored pairs and adds speed with the same logic. The cost is the obvious one: 50% of raw capacity. For OS volumes, databases, VM datastores — anything where latency and rebuild simplicity matter more than capacity — mirrors are the default that rarely disappoints.
RAID5: retired on large drives, and why
Single parity survives one failure. The problem is what happens during the rebuild: every remaining drive must be read end to end, and with 10–16TB disks that window stretches to a day or more of maximum-stress reads on aging siblings of the drive that just died. A second failure — or a single unrecoverable read error at the wrong moment — ends the array. On small fast SSDs, RAID5 still has a case; on big spinning drives, it is a bet that gets worse every drive generation.
RAID6: the capacity workhorse
Double parity survives any two failures, which is precisely what makes large-drive arrays sane: one dead drive plus one read error is Tuesday, not a disaster. The costs are a heavier write penalty (each random write touches data plus two parity updates) and the same long rebuild windows — you are just protected while they run. RAID6 (and ZFS's RAIDZ2) is the standard answer for bulk file, media and backup pools.
Layouts by bay count
2 bays: RAID1, done. 4: RAID10 for speed, RAID6 only if capacity is desperate. 8: the classic split — two-drive RAID1 for OS/apps plus six-drive RAID6 for data; or all-RAID10 for an all-flash VM store. 12: two six-wide RAID6 groups rather than one twelve-wide — halves rebuild exposure for a small capacity cost. Always add a hot spare when the chassis has a free bay and the array is parity-based.
Three closing rules
Battery/flash-backed cache on the controller is what makes parity writes tolerable — verify it works. Patrol reads / scrubs exist to find the latent errors that kill rebuilds; leave them on. And RAID is availability, not backup: it survives dead drives, not deletion, ransomware or a controller writing nonsense. The array keeps you running; the backup brings you back.