RAID is the one storage decision people copy from the last server without thinking, and it is also the one most likely to lose data when a drive fails. The right level is not a matter of taste - it falls out of three numbers: how many drives you have, how big each one is, and how write-heavy the workload is. Pick the wrong level on large modern drives and a single failure can turn into a multi-day rebuild with a real chance of a second failure mid-way. This is the framework our engineers use to choose a RAID level for UK customers.
What RAID actually buys you (and what it does not)
RAID protects against drive failure and, in some levels, improves throughput. It is not a backup - it will happily replicate a ransomware encryption or an accidental delete across every member drive instantly. Decide RAID for availability, then back up separately for recovery. With that clear, the choice between RAID 1, 5, 6 and 10 comes down to a trade between usable capacity, how many drives can fail, how long a rebuild takes, and the write penalty you pay for parity.
Every parity level carries a write penalty: each logical write turns into multiple physical operations because the controller has to read old data, read old parity, and write both back. That penalty is invisible on a lightly-loaded file server and brutally obvious on a busy database. Match the level to the workload, not to a habit.
RAID 1 and RAID 10: mirror first, ask questions later
RAID 1 mirrors two drives - simple, fast to rebuild, and the obvious choice for a boot pair or a small two-drive server. You lose half your raw capacity but gain a trivial rebuild: the controller just copies one surviving drive to its replacement, with no parity maths and minimal performance hit.
RAID 10 stripes across mirrored pairs and is the default for transactional databases and busy virtualisation. It has no parity write penalty, excellent random write performance, and fast rebuilds because only one drive in a pair is copied. The cost is capacity - you still only get half your raw drives - but for latency-sensitive write workloads on NVMe it is almost always the right answer.
RAID 5 and RAID 6: capacity with a rebuild-risk tail
RAID 5 gives you single-drive fault tolerance with only one drive's worth of capacity overhead, which is why it is so tempting on larger arrays. The problem is the rebuild. When a drive fails, every remaining drive must be read in full to reconstruct the lost data onto the replacement, and on today's 16-22TB drives that read can run for many hours or days under load. If a second drive develops an unrecoverable read error during that window, the array is lost. On large-capacity drives RAID 5 has effectively aged out for primary data.
RAID 6 adds a second parity block, surviving two simultaneous drive failures, which is what makes it the sane default for large near-line arrays. It pays a heavier write penalty than RAID 5 but tolerates a failure during a rebuild - exactly the scenario that kills RAID 5 on big disks. For bulk capacity, backup repositories and media stores on high-capacity drives, RAID 6 is the level that respects rebuild risk.
- •RAID 1: boot pairs and small servers - trivial rebuild, half capacity
- •RAID 10: databases and busy VMs - no parity penalty, best random writes
- •RAID 5: small fast-drive arrays only - one failure tolerance, risky rebuild on big disks
- •RAID 6: large near-line/backup arrays - survives a second failure mid-rebuild
The write penalty matters more on NVMe than people expect
It is easy to assume NVMe is so fast that parity overhead disappears. In practice the opposite tension appears: a traditional RAID controller can become the bottleneck in front of multiple Gen4 or Gen5 NVMe drives, and the parity write penalty eats into the very low latency you paid for. For write-heavy NVMe workloads, RAID 10 (or software/HBA-based approaches that pass the drives through) usually beats parity RAID on a hardware controller.
If you are building an all-NVMe tier, think carefully about whether you want a RAID controller in the path at all, or a straight-through host bus adapter feeding software-defined storage. Choose the drives deliberately too - endurance and mixed-use rating matter as much as the RAID level - using our SSD and NVMe guidance.
Putting it together: a level for each tier
Most real servers end up with more than one RAID level: a mirrored boot device, RAID 10 for the database or hypervisor datastore, and RAID 6 for bulk capacity. Decide each tier from drive count, drive size and write profile rather than applying one level to the whole box. Keep a hot spare on any parity array so a rebuild can start automatically instead of waiting on a callout.
When you have settled the levels, build the exact controller, drive and cage combination in our configurator and let us validate it against the chassis. Start from our server configuration service, and read the broader method in how to spec a server in 2026.