Choosing RAID for a Bare-Metal Server: RAID 0 vs 1 vs 5 vs 10

When you rent a bare-metal (dedicated) server with several drives, you usually have to pick how those drives are combined into a RAID array. Get it right and a single dead disk costs you nothing while reads and writes stay fast. Get it wrong and you either waste capacity or lose everything the moment one drive fails. This guide helps you decide between RAID 0, 1, 5, and 10.

How each level works

RAID 0 (striping)

Data is split into chunks and spread across every disk for parallel I/O. You get the best performance and 100% usable capacity, but zero redundancy—lose any single drive and the entire array is gone. Reserve it for scratch space or data you can rebuild at will.

RAID 1 (mirroring)

Two drives hold identical copies; if one dies, the other keeps serving. Strong redundancy and solid read speed, at the cost of half your capacity and every write landing twice. A great fit for OS drives and smaller, stability-sensitive volumes.

RAID 5 (distributed parity)

Three or more drives store data plus one drive's worth of parity spread across the array, so it survives any one drive failure. Usable capacity is (N-1)/N—more economical than mirroring. The trade-offs: a write penalty from parity math, and degraded performance during a rebuild where a second failure kills the array. Rebuild risk grows sharply with large-capacity disks.

RAID 10 (mirror + stripe)

Mirror first, then stripe—combining RAID 1's redundancy with RAID 0's speed. It needs at least four drives and gives you 50% usable capacity, but delivers excellent random I/O, fast rebuilds, and strong fault tolerance. The safe choice for demanding workloads.

At a glance

| Level | Min. drives | Fault tolerance | Usable capacity | Relative performance | Typical use | |-------|------------|-----------------|-----------------|----------------------|-------------| | RAID 0 | 2 | None | 100% | Highest | Scratch / rebuildable data | | RAID 1 | 2 | 1 drive | 50% | Good reads, average writes | OS drive, general | | RAID 5 | 3 | 1 drive | (N-1)/N | Good reads, write penalty | General, read-heavy | | RAID 10 | 4 | 1 per mirror | 50% | Best random I/O | Databases, high I/O |

How to choose

• Databases, virtualization, heavy random I/O: reach for RAID 10—best performance and the safest rebuilds.
• General-purpose, web, and file serving: on a budget, pick RAID 1 (few drives) or RAID 5 (more drives, read-heavy workloads where saving capacity matters).
• Purely temporary, always rebuildable data: only then consider RAID 0—and think twice in production.

Hardware RAID card vs. software RAID

• Hardware RAID card: a dedicated controller handles parity and carries its own cache (pair it with a BBU or capacitor for power-loss protection). It's transparent to the OS and consistently fast, but ties you to the vendor—swap the card and the array may become unreadable.
• Software RAID (Linux mdadm): no special card required. It's flexible, portable across machines, and transparent to troubleshoot, using a little CPU that's negligible on modern hardware. For many setups it's the pragmatic choice.

When ordering a server, ask your provider which type is in use and whether power-loss protection is included.

RAID is not a backup

This is the one to remember: RAID only protects against physical drive failure. It does nothing for accidental deletion, ransomware, file corruption, or a whole-machine disaster—those mistakes replicate to every disk instantly. RAID buys availability, not data safety. Always keep independent, off-site, regular backups—and verify you can actually restore them.

Summary

RAID 0 trades safety for raw speed, RAID 1 is simple and dependable, RAID 5 saves capacity for read-heavy work, and RAID 10 is the all-around winner under load. Match the level to the workload—RAID 10 for databases and high I/O, RAID 1 or 5 for general use—and weigh a hardware card against mdadm on its own merits. Above all: no RAID level, ever, replaces a backup you've proven you can restore.