When a disk fails in a recording server, the real question isn’t “does recording stop?” but “can we survive a second failure until the rebuild finishes?” Rebuild time depends on capacity, effective speed, recording load and storage architecture; in the worked example below, it comes to about 44–89 hours for a 16 TB disk. The wrong RAID level turns a single disk failure into a risk of loss lasting that entire period. In this guide we compare RAID 5, 6, 10, nested levels and non-RAID approaches for NVR and VMS recording servers in terms of capacity, write performance and rebuild risk.
How video load strains storage
In enterprise video surveillance systems, the recording server writes all day and rarely reads. A single camera stream is a sequential write. But when 50–200 streams write to different files at the same time, the disk sees this traffic as a semi-random pattern. Reads, apart from live viewing, arrive in short, intense bursts during playback, incident review and export. If restoring video recordings is part of the facility’s continuity plan, define the test order and retention policy within enterprise backup and disaster recovery.
This pattern has three consequences for RAID selection:
- The write penalty is decisive. In levels using parity, every small write requires a read–modify–write cycle. In the classic calculation, one logical write corresponds to 4 physical I/Os in RAID 5, 6 in RAID 6 and 2 in RAID 10. VMSs usually write in large blocks, and the controller cache can accumulate full-stripe writes, so the real penalty is lower. On a system with no cache or with the cache disabled, however, these figures are felt in full.
- The write cache must be protected. RAID 5/6 performance relies heavily on write-back cache. If you enable write-back on a controller without battery- or flash-backed protection, the array can be left inconsistent after a power cut; if you disable it, write performance drops. A common situation is a controller that has silently switched to write-through because its cache battery has been faulty for months.
- Degraded mode is a real operating condition. Recording doesn’t stop after a disk failure; the array keeps carrying the same load until the rebuild finishes. Design for the degraded and rebuild state, not the normal state.
RAID level comparison table
The table below is a worked example for a recording server with 12 × 16 TB disks. Usable capacity is the raw value; free space reserved by the file system and the VMS is deducted separately. Disk manufacturers state 16 TB in decimal (10¹² bytes), and the operating system shows it as about 14.55 TiB. Don’t let this difference surprise you in capacity planning.
| Level | Usable (12 × 16 TB) | Fault tolerance | Write penalty | Notes for video |
|---|---|---|---|---|
| RAID 5 | 176 TB | 1 disk | 4 | High rebuild risk with large disks |
| RAID 6 | 160 TB | Any 2 disks | 6 | The default choice at medium scale |
| RAID 10 | 96 TB | 1 disk per mirror | 2 | High capacity cost, fast rebuild |
| RAID 50 (2×6) | 160 TB | 1 disk per group | 4 | Load is spread; risk close to RAID 5 |
| RAID 60 (2×6) | 128 TB | 2 disks per group | 6 | A balanced choice for large arrays |
RAID 10’s real strength is low latency and random I/O performance; video recording often doesn’t need this, while half the capacity goes to mirroring. Consider RAID 10 for small, read-heavy systems (many simultaneous playbacks, analytics reprocessing) or for VMS database and cache disks. For the recording pool itself, the real choice is mostly between RAID 6 and RAID 60.
Rebuild time and risk with large disks
In classic RAID, a rebuild means writing the new disk across its full capacity, and its speed is limited by a single disk’s sustained write speed. While recording load continues, the controller can’t give the rebuild all of its bandwidth.
Worked example: If a 16 TB disk is rebuilt at an average of 100 MB/s under recording load, the time is 16,000,000 MB ÷ 100 MB/s = 160,000 seconds, or about 44 hours. If the effective speed drops to 50 MB/s because of load, the time rises to about 89 hours, and for a 24 TB disk at the same speed to about 133 hours. In RAID 5, the array is unprotected against a second failure for this entire period. In this example, for a 24 TB disk there’s no tolerance for a second failure for about five and a half days. The calculation assumes a constant effective speed and rewriting across full capacity; in architectures that copy only used space or use distributed spare space, the time is different.
RAID is fault tolerance, not backup: it doesn’t protect recordings against deletion, fire, theft, controller failure or loop recording overwriting them. An independent copy (secondary recording, archive or incident export) is planned separately for critical recordings.
Three factors increase the risk during a rebuild:
- Disks from the same batch: Disks bought on the same day, put into service at the same hour and carrying the same load wear out at similar times. The first failure is often a harbinger of the second.
- Unrecoverable read errors (URE): A rebuild requires reading all remaining disks from start to finish. Datasheets usually give the URE rate as 1 in 10¹⁴ or 10¹⁵ bits; 10¹⁴ bits is about 12.5 TB. This is an upper-bound specification, and the real rate is often better. Still, when the data read in a RAID 5 rebuild reaches this order of magnitude, a single unreadable sector means the affected stripe can’t be recovered.
- Increased load: Rebuild reads, recording writes and heavy playback demand after an incident arrive at the same time. A failure is often noticed exactly on the days when people want to review recordings.
That’s why we don’t recommend RAID 5 for the recording pool with disks of 8 TB and above; use RAID 6 or RAID 60. Also check the controller’s rebuild priority setting: high priority shortens the rebuild but lowers recording performance. The right value is the highest speed you can reach in the acceptance test in the degraded state without frame loss. It’s set by measuring, not by guessing.
On recording servers using software RAID (Linux md), you monitor rebuild speed limits and status like this:
# Array status and rebuild progress
cat /proc/mdstat
mdadm --detail /dev/md0
# Rebuild speed limits (in KB/s)
cat /proc/sys/dev/raid/speed_limit_min
cat /proc/sys/dev/raid/speed_limit_max
# Example of temporarily raising the minimum when recording load is low
echo 100000 > /proc/sys/dev/raid/speed_limit_min
On hardware RAID controllers, the same information is read from the manufacturer’s management tool: rebuild rate, and patrol read and consistency check scheduling. If patrol reads and consistency checks run regularly, unreadable sectors are found while the array is healthy, not during a rebuild. That’s also the first place to look in the console: when did the last patrol read complete?
Array width and number of disks
How many disks you put in a single RAID 6 group is a balance between capacity efficiency and risk. As the number of disks in the group grows, the parity share falls and usable capacity rises. But the amount of data to be read during a rebuild and the number of disks that can fail at the same time also grow.
- 4–6 disks: Small NVRs. With four disks, RAID 6 and RAID 10 give the same capacity; RAID 6 is safer because it survives the loss of any two disks.
- 8–12 disks: A reasonable range for a single RAID 6 group. The parity share is acceptable.
- More than 12 disks: Building several RAID 6 groups (RAID 60, or separate volumes on the VMS side) is usually better than one wide group. One group’s rebuild affects the other groups less.
There’s a decision on the VMS side too. If you define a separate recording volume per group instead of a few large volumes, you can distribute cameras across volumes; when a group is degraded, only the cameras writing to that group are affected. For the capacity calculation, use the method in the how many days does an NVR record and how is storage calculated guide together with the usable capacity values here.
Hot spare
A hot spare is a disk that waits idle in the array and that the controller automatically takes into a rebuild at the moment of failure. In recording systems, days can pass between a failure being noticed and someone responding; a hot spare directly shortens the period spent unprotected.
- Global hot spare: Serves several groups and uses disk slots efficiently. Preferred in large arrays.
- Dedicated hot spare: Assigned to a single group; the critical group always has a spare.
- The hot spare should be of the same or larger capacity and the same class as the disks in the group.
- A hot spare doesn’t replace monitoring and alarms. A notification should also arrive when a rebuild starts, and the failed disk should be replaced the same week.
Non-RAID recording approaches
Not every recording system needs RAID. In some scenarios, a different protection arrangement makes more sense than RAID’s cost and complexity:
- Independent disks (JBOD-like use): Some NVRs and VMSs use each disk as a separate recording volume and distribute cameras across disks or disk groups. When a disk fails, only the recordings written to that disk are lost; other cameras aren’t affected. Capacity efficiency is highest; recordings on the failed disk, however, don’t come back.
- Redundant recording: Critical cameras are recorded simultaneously to two separate recording servers or NVRs. This covers disk, server and storage failures alike; it increases network and licensing costs.
- Supplementing with edge recording: The camera records to its SD card continuously or only while the connection is down, and the gap on the recording server is filled in afterwards. Details are in the edge recording and ANR guide.
- Scalable storage using erasure coding: In large central installations, data is split and written across several nodes; disk and node failures are tolerated together. Be sure to verify that the VMS supports this storage type for continuous writing.
Scenarios in which the recording server goes completely offline are outside RAID’s scope. We covered this subject in the VMS failover and N+1 design guide.
Selection checklist
- Have the total write bandwidth (the sum of all camera bitrates) and the simultaneous playback requirement been documented?
- If disk capacity is 8 TB or more, has RAID 5 been removed from the recording pool?
- For arrays larger than 12 disks, have multiple RAID 6 groups or RAID 60 been considered?
- Does the controller have battery- or flash-backed write-back cache, and is the cache protection healthy?
- Has at least one hot spare been defined, and was global or dedicated chosen deliberately?
- Are patrol reads and consistency checks scheduled?
- Do disk failure and rebuild start alarms reach a recipient?
- In the acceptance test, was a disk pulled and frame loss measured in the degraded and rebuild states?
- Were the disks chosen from a class suitable for video load? (Details: surveillance disk selection)
If you’re experiencing disk failures or insufficient capacity in an existing system, we can assess it under our NVR disk failure and capacity upgrade service; for designing the protection scheme and capacity together in a new project, under our recording storage and capacity planning service. The official reference for software RAID behavior on Linux-based recording servers is the Linux kernel md documentation.
Frequently asked questions
With 4 disks in a small NVR, which should be chosen: RAID 5, RAID 6 or RAID 10?
With four disks, RAID 10 and RAID 6 give the same usable capacity (two disks). RAID 6 survives any two disk failures; RAID 10 leads in write performance and rebuild speed. We don’t recommend four-disk RAID 5 with disks of 8 TB and above, because it’s left unprotected against a second failure during the rebuild.
Does RAID replace backups of recordings?
No. RAID only preserves availability against disk failure. It doesn’t protect recordings against accidental deletion, ransomware, controller failure or theft of the device. Set up a separate archive and backup arrangement for incident recordings that must be preserved.
Does recording stop during a rebuild?
Normally not; the array keeps writing in degraded mode. However, parity calculation and rebuild read load lower write performance, and if load is at the limit, dropped frames or recording gaps can occur. That’s why you should measure recording performance under rebuild in the acceptance test.
Does using SSDs change the RAID choice?
SSDs are rarely used for video archives because of capacity cost. If they are, rebuild time shortens markedly; the protection level still has to be chosen according to fault tolerance and capacity needs.