In plain words
Storage is the warehouse for your data. The disks are the shelves; RAID means storing each box twice (or with a card that lets you rebuild it) so that a collapsing shelf loses nothing; the file system is the storage plan; the NAS is the shared warehouse everyone reaches over the network. One rule to keep in mind right now: storing things twice in the same warehouse doesn't protect against a fire. RAID is not a backup.
1. Disks: HDD or SSD
| Hard drive (HDD) | SATA SSD | NVMe SSD | |
|---|---|---|---|
| Principle | Spinning magnetic platters, moving read head | Flash memory, SATA interface | Flash memory, plugged directly into the PCIe bus |
| Typical sequential speed | 150 to 280 MB/s | About 550 MB/s (the SATA limit) | 3,500 MB/s (PCIe 3.0), 7,000 MB/s (PCIe 4.0), over 12,000 MB/s (PCIe 5.0) |
| Random access | Slow (the head has to move) | Very fast | The fastest |
| Capacity | Up to over 30 TB | Up to a few TB | Up to a few TB (more for servers) |
| Wear | Mechanical: vibration, shocks, running hours | Limited number of writes (TBW) | Same as SATA SSD |
| Noise, power | Audible, uses more power | Silent, frugal | Silent, frugal, gets hot under load |
| Ideal use | Large volumes: films, photos, backups | System, VMs, replacing an old hard drive | System, VMs, databases, cache |
SSD endurance: TBW and DWPD
A flash memory cell can only take a limited number of writes. Manufacturers express this as:
- TBW (terabytes written): the total amount of data that can be written before wear-out, for example 600 TB for a 1 TB consumer SSD;
- DWPD (drive writes per day): how many times the whole drive can be rewritten every day during the warranty. 0.3 DWPD suits a workstation; a cache or database SSD aims for 1 to 3 DWPD.
For a workstation, endurance is almost never a problem. For a NAS cache or a hypervisor with ZFS, it matters: choose high-endurance models.
CMR or SMR: the NAS drive trap
Hard drives write their tracks in two ways:
| CMR (conventional recording) | SMR (shingled recording) | |
|---|---|---|
| Principle | Tracks side by side, independent | Tracks that overlap, like roof tiles, to gain density |
| Writes | Steady | Can collapse during sustained writes: neighbouring tracks must be rewritten |
| In RAID | Suitable | Not recommended: rebuilding an array can become extremely slow, or even fail with ZFS |
Check the technology before buying
The manufacturer states the technology in the datasheet, not always on the box. A well-known example: at Western Digital, plain WD Red drives are SMR, while WD Red Plus and WD Red Pro are CMR. Seagate's NAS ranges (IronWolf, IronWolf Pro) and Toshiba's (N300 series) are CMR. For a NAS or a server in RAID, buy CMR.
2. RAID: combining several disks
RAID groups several disks into a single volume, to gain fault tolerance, performance, or both. It protects against one specific thing: a disk failure. It doesn't protect against accidental deletion, ransomware, fire or a handling mistake, which hit all the disks at once.
| Level | Principle | Min. disks | Usable capacity (n disks of size D) | Failures tolerated | Use |
|---|---|---|---|---|---|
| RAID 0 | Data striped across all the disks | 2 | n × D | None: lose one disk, lose everything | Temporary data, video editing; never for data you want to keep |
| RAID 1 | Mirror: each disk holds everything | 2 | D | n − 1 | System, small 2-bay NAS |
| RAID 5 | Stripes + distributed parity | 3 | (n − 1) × D | 1 | Small volumes; risky with large disks (see below) |
| RAID 6 | Stripes + two parities | 4 | (n − 2) × D | 2 | NAS with 4 to 8 disks or more, large disks |
| RAID 10 | Mirrors combined into stripes | 4 | n × D / 2 | 1 per mirror (at least 1, sometimes more) | Databases, VMs: performance and safety |
Concrete examples with 4 disks of 8 TB
| Level | Usable capacity | Failures tolerated |
|---|---|---|
| RAID 0 | 32 TB | 0 |
| RAID 5 | 24 TB | 1 |
| RAID 6 | 16 TB | 2 |
| RAID 10 | 16 TB | 1 (2 if they hit different mirrors) |
Why RAID 5 is not recommended with large disks
When a disk dies, the array has to be rebuilt: all the remaining disks are read in full to recalculate the lost data. With disks of 10 TB and more, this rebuild often takes more than a day, under heavy load. Meanwhile, the array can no longer tolerate any failure: one read error or a second failing disk (often from the same batch, the same age) and everything is lost. Beyond 4 disks or with disks of several TB, prefer RAID 6 (or RAIDZ2).
NAS- and ZFS-specific variants
| Technology | Where | Special feature |
|---|---|---|
| SHR and SHR-2 (Synology Hybrid RAID) | Synology NAS | Equivalents of RAID 5 and RAID 6 that make the best use of disks of different sizes |
| RAIDZ1, RAIDZ2, RAIDZ3 | ZFS: TrueNAS, Proxmox VE, QNAP (QuTS hero) | Equivalents of RAID 5, 6 and "triple parity"; ZFS also checks the integrity of every block |
| ZFS mirror | ZFS | Equivalent of RAID 1 (or RAID 10 by combining several mirrors) |
| Unraid parity | Unraid | Independent disks of different sizes, protected by one or two parity disks; each disk remains readable on its own |
Hardware or software RAID?
| Hardware RAID | Software RAID | |
|---|---|---|
| Managed by | A dedicated controller card | The system: mdadm (Linux), ZFS, Btrfs, Storage Spaces (Windows) |
| Advantage | Transparent to the system, battery-protected cache on high-end cards | No dependence on a card: the disks can be read on any machine |
| Drawback | If the card dies, you need a compatible model to read the disks | Uses a little CPU (negligible today) |
| Trend | Classic enterprise servers | NAS, homelab, ZFS: the norm |
The "RAID" offered by the BIOS of consumer motherboards (Intel RST, AMD RAIDXpert) is a mix of the two, with the advantages of neither: avoid it on Linux and for a NAS.
3. DAS, NAS, SAN: where are the disks?
| DAS | NAS | SAN | |
|---|---|---|---|
| Meaning | Direct Attached Storage | Network Attached Storage | Storage Area Network |
| Principle | Disks plugged directly into a machine (internal, USB, Thunderbolt) | A device on the network that shares files | A dedicated network that provides raw disks (blocks) to servers |
| Access | A single machine | Several computers, over the local network | Servers, which create their own file system on them |
| Protocols | SATA, USB, Thunderbolt, SAS | SMB, NFS, AFP (old), WebDAV | iSCSI, Fibre Channel, NVMe-oF |
| Example | External backup drive | Synology, QNAP, TrueNAS at home | Enterprise storage array for a virtualisation cluster |
A modern NAS can do both: share files (SMB, NFS) and present volumes over iSCSI, like a small SAN.
4. File systems
The file system organises the data on the disk: names, folders, permissions, the location of each block.
| System | Platforms | Strengths | Limits |
|---|---|---|---|
| NTFS | Windows (read and written by Linux and, read-only, by macOS) | Fine-grained permissions, journaling, the Windows standard | Poorly supported outside Windows |
| ReFS | Windows Server, Windows 11 (some editions) | Resistance to corruption, large volumes | Can't boot from it on a regular PC, features vary by edition |
| exFAT | Windows, macOS, Linux | Read and written everywhere: ideal for a USB stick or a shared external drive | No journaling: fragile if unplugged abruptly |
| APFS | macOS | Snapshots, encryption, optimised for SSD | Apple-only |
| ext4 | Linux | Mature, fast, very reliable | No snapshots or data verification |
| XFS | Linux | Excellent with large files and large volumes | Can't be shrunk |
| Btrfs | Linux, Synology NAS and others | Snapshots, checksums, compression | Its native RAID 5 and 6 modes are still not recommended; Synology uses mdadm underneath |
| ZFS | Linux, FreeBSD, TrueNAS, Proxmox VE | Built-in disk and RAID management, checksums on every block, self-healing, snapshots, compression, replication | Memory-hungry; a pool is less flexible to change |
Why checksums matter
A disk can, rarely, return corrupted data without reporting an error. ext4 and NTFS don't notice: the damaged file is served, then backed up as it is. ZFS and Btrfs compute a fingerprint for each block: on reading, corrupted data is detected, and repaired automatically from the healthy copy if the volume is redundant. A periodic scrub rereads the whole volume to find these errors before they pile up: schedule one a month.
Snapshots
Btrfs, ZFS and APFS can freeze the state of a volume in a fraction of a second, without copying the data. After a deletion or ransomware, you go back to yesterday's snapshot. On a NAS, schedule regular snapshots (every hour, kept for a few days): it's the best defence against a handling mistake. It's still not a backup: snapshots live on the same disks.
5. Sharing protocols
| Protocol | For whom | Notes |
|---|---|---|
| SMB (Windows sharing, also called CIFS) | Windows, macOS, Linux | The universal standard. Use SMB 3 and disable SMB 1, old and vulnerable |
| NFS | Linux, hypervisors (Proxmox VE, ESXi) | Very efficient between Linux machines; NFS v4 adds Kerberos authentication |
| iSCSI | Servers, hypervisors | The NAS provides a raw disk: the client machine formats it itself. Never mount the same iSCSI volume on two machines without a cluster file system |
| WebDAV | Web access, mobile apps | Handy across the internet, to be protected with HTTPS |
| S3 | Backups, applications | "Object" storage; many backup programs can write to S3 |
| AFP | Old Macs | Dropped by Apple: use SMB |
6. Real-world speeds: the network is often the limit
| Link | Theoretical speed | Measured usable speed |
|---|---|---|
| 1 Gbit/s Ethernet | 125 MB/s | 110 to 118 MB/s |
| 2.5 Gbit/s Ethernet | 312 MB/s | About 280 MB/s |
| 10 Gbit/s Ethernet | 1,250 MB/s | 1,000 to 1,150 MB/s |
| Wi-Fi 6 | Variable | Often 50 to 120 MB/s in good conditions |
| USB 3.2 Gen 1 (5 Gbit/s) | 625 MB/s | About 400 MB/s |
In practice: a single hard drive already saturates a 1 Gbit/s link. Putting SSDs in a NAS connected at 1 Gbit/s doesn't speed up file copies; moving to 2.5 Gbit/s (network card and switch) is often the best upgrade. To measure the network's real speed, use iperf3 (see the software catalogue).
7. ECC memory
ECC memory (Error-Correcting Code) detects and corrects bit errors in memory. It's the norm on servers. For a NAS holding important data, especially with ZFS, it is recommended but not mandatory: ZFS works without it, just with a residual risk that a memory error corrupts data before it's written. It needs a compatible processor and motherboard.
8. The golden rules
- RAID is not a backup. Keep at least one copy of the data on another medium, including one off-site: see the 3-2-1 rule.
- For a NAS: CMR disks, preferably from a NAS range, ideally from different batches so they don't die together.
- RAID 1 for 2 disks, RAID 6 (or SHR-2, RAIDZ2) from 4 to 5 disks or with large disks.
- Monitor the disks' SMART status and enable the NAS's email alerts: a degrading disk gets replaced before it fails.
- Schedule a monthly scrub with ZFS or Btrfs, and regular snapshots.
- Keep a spare disk to rebuild quickly.
Key points
- HDD for capacity, SSD for speed; check that disks meant for RAID use CMR technology.
- RAID protects against a disk failure, nothing else; RAID 5 is risky with large disks, prefer RAID 6.
- NAS = files shared over the network (SMB, NFS); SAN = raw disks for servers (iSCSI).
- ZFS and Btrfs detect silent corruption and offer snapshots.
- The network is often more of a limit than the disks: 1 Gbit/s ≈ 115 MB/s.