RAID 5 vs 6 vs 10: Capacity, Write Penalty, and Rebuild Risk on the Same Eight Drives
The same eight 8 TB drives give you 56 TB, 48 TB, or 32 TB depending on the level, and between 576 and 1,108 write-weighted IOPS. Here is every trade-off costed, including the rebuild window that makes RAID 5 dangerous on large disks.
RAID 5 vs 6 vs 10
Three ways to arrange the same disks, trading capacity against write speed against how many failures you survive. There is no best one; there is a correct one for a given workload and drive size, and the drive size matters more than most people expect.
This article uses one hardware configuration throughout, eight 8 TB drives, and works out capacity, throughput, and rebuild exposure for each level so the numbers are directly comparable.
Try the RAID calculatorEnter drive count and size for any level to get usable capacity, fault tolerance, and storage efficiency.What each level actually does
Swipe sideways to compare columns.
| RAID 5 | RAID 6 | RAID 10 | |
|---|---|---|---|
| Mechanism | Striping with one parity block | Striping with two parity blocks | Mirrored pairs, striped |
| Minimum drives | 3 | 4 | 4, and always even |
| Usable capacity | (n − 1) × drive | (n − 2) × drive | n ÷ 2 × drive |
| Guaranteed failures survived | 1 | 2 | 1 |
| Best case failures survived | 1 | 2 | n ÷ 2, if no two are partners |
| Write penalty | 4 | 6 | 2 |
| Rebuild reads | Every remaining drive | Every remaining drive | One drive, the mirror partner |
The write penalty row drives most of the performance difference. A single small write to a RAID 5 array is four physical operations: read the old data block, read the old parity block, write the new data, write the new parity. RAID 6 maintains two independent parity blocks, so it is six. RAID 10 simply writes to both halves of a mirror, so it is two.
Capacity on eight 8 TB drives
Swipe sideways to compare columns.
| Level | Usable | Efficiency | Capacity given up |
|---|---|---|---|
| RAID 5 | 56 TB | 87.5% | 8 TB |
| RAID 6 | 48 TB | 75.0% | 16 TB |
| RAID 10 | 32 TB | 50.0% | 32 TB |
Efficiency improves with drive count for the parity levels and never does for RAID 10. Across twenty-four drives, RAID 5 reaches 95.8% and RAID 6 reaches 91.7%, while RAID 10 is still exactly 50%. That is the argument for parity on large archival arrays, and it is a real one.
Throughput, with the write penalty applied
Random I/O is where the levels diverge. Take eight drives capable of 180 IOPS each, so 1,440 raw IOPS, running a workload that is 70% reads and 30% writes.
RAID 10 delivers 46% more usable I/O than RAID 5 and 92% more than RAID 6, on identical hardware, for half the capacity. Now change the workload to 70% writes and the gap widens sharply.
Swipe sideways to compare columns.
| Read / write mix | RAID 5 | RAID 6 | RAID 10 |
|---|---|---|---|
| 100% read | 1,440 | 1,440 | 1,440 |
| 90 / 10 | 1,108 | 960 | 1,309 |
| 70 / 30 | 758 | 576 | 1,108 |
| 50 / 50 | 576 | 411 | 960 |
| 30 / 70 | 465 | 320 | 847 |
The rebuild window, which is the real argument
Capacity and speed are visible on a spreadsheet. Rebuild risk is not, and it is what has pushed RAID 5 out of favour for large drives.
When a parity array loses a drive, rebuilding the replacement requires reading every remaining drive in full. On the eight-drive RAID 5, that is 56 TB of reading. At a sustained 80 MB/s, which is realistic for an array still serving production traffic, the rebuild is gated by how fast one drive can be written, about 28 hours at best, and multi-day rebuilds on large arrays are routine.
Swipe sideways to compare columns.
| RAID 5 | RAID 6 | RAID 10 | |
|---|---|---|---|
| Data read during rebuild | 56 TB across 7 survivors | 56 TB across 7 survivors | 8 TB, one partner |
| State during rebuild | Zero redundancy | Still tolerates one more failure | That pair unprotected, rest fine |
| A second failure means | Total array loss | Rebuild continues | Loss only if it is the partner |
| Performance during rebuild | Severely degraded | Severely degraded | Mildly degraded |
| Rough rebuild time at 80 MB/s per drive | ~28 hours | ~28 hours | ~28 hours, at far lower risk |
The row that matters is the second one. A RAID 5 array in rebuild has no redundancy at all, for over a day, while working the surviving drives harder than they have been worked since the array was built. Drives bought in one batch and aged identically are precisely the population most likely to fail together.
Unrecoverable read errors and why drive size changed the maths
Drive datasheets quote an unrecoverable read error rate, typically one error per 10¹⁴ bits for consumer SATA and one per 10¹⁵ or 10¹⁶ for enterprise drives. One error per 10¹⁴ bits is one per 12.5 TB read.
A RAID 5 rebuild on the array above reads 56 TB. Against a 12.5 TB error budget, that is roughly four and a half times the rated interval, which makes hitting at least one error during a rebuild more likely than not. In RAID 5 with no redundancy left, an unrecoverable read error during rebuild can cost the array or at minimum the affected data.
Swipe sideways to compare columns.
| Array | Data read | Multiples of a 12.5 TB error budget |
|---|---|---|
| 4 × 2 TB | 6 TB | 0.5 |
| 6 × 4 TB | 20 TB | 1.6 |
| 8 × 8 TB | 56 TB | 4.5 |
| 12 × 16 TB | 176 TB | 14.1 |
Enterprise drives rated at one error per 10¹⁵ bits push the budget to 125 TB and make the arithmetic far more comfortable. Check the datasheet rather than assuming, because the rating varies by an order of magnitude across drives that look identical on a price list.
Choosing
Swipe sideways to compare columns.
| Situation | Level | Reason |
|---|---|---|
| Database or VM host | RAID 10 | Write penalty of 2 and fast, low-risk rebuilds |
| Backup target or media archive | RAID 6 | Capacity efficiency with two-failure tolerance |
| Large drives, 8 TB and up | RAID 6 or 10 | RAID 5 rebuild exposure is too long |
| Small array of small enterprise drives | RAID 5 is defensible | Short rebuild, high error budget |
| Read-mostly file server | RAID 6 | Reads carry no penalty on parity levels |
| Anything without a tested backup | None of them | RAID is availability, not backup |
What these numbers do not tell you
- RAID is not backup. Every level above protects against drive failure and none of them protects against deletion, ransomware, filesystem corruption, controller failure, or fire.
- The IOPS model is a simplification. Controller cache, write coalescing, and full-stripe writes can reduce the effective write penalty substantially on sequential workloads.
- Solid state drives change the picture. Their IOPS are high enough that the write penalty rarely constrains throughput, though it still consumes write endurance, which is a real cost on parity levels.
- The unrecoverable read error rate is a specification, not a measured failure probability. Treating it as an exact probability overstates the precision; treating it as irrelevant understates the risk.
- Modern controllers and filesystems differ. ZFS RAID-Z and similar implementations avoid the parity read-modify-write cycle and do not have the same write penalty as classic RAID 5.
- Rebuild times assume a healthy array under moderate load. A busy array rebuilds far more slowly, and some controllers deliberately throttle rebuilds to preserve foreground performance.
- Nothing here accounts for the cost of the drives themselves, which is often the deciding factor and which favours parity strongly.
Is RAID 5 obsolete?
Not obsolete, but out of its range for large drives. On a small array of enterprise disks with a short rebuild window and a tested backup, it is still reasonable. On 8 TB or larger consumer drives, the rebuild reads enough data to make an unrecoverable read error likely, and RAID 6 costs one more drive.
How many drives can RAID 10 lose?
At least one, guaranteed. Potentially half the array, if no two failures land in the same mirrored pair. On an eight-drive array the second failure has a one in seven chance of being the wrong drive, so treat the guarantee as one and the rest as luck.
Why is RAID 6 slower than RAID 5 on writes?
It maintains a second, independently computed parity block. Every small write becomes six operations rather than four: read data, read both parity blocks, write data, write both parity blocks. That is a 50% higher penalty for one extra drive of fault tolerance.
Does RAID 10 need double the drives for the same capacity?
Yes, exactly double. To reach the 56 TB that eight drives give under RAID 5, RAID 10 needs fourteen. That is the price for the write penalty of 2 and the single-drive rebuild, and for write-heavy workloads it is usually worth paying.
Should I use a hot spare?
On parity arrays, yes. The rebuild starts immediately rather than waiting for someone to notice the alert and walk to the rack, and on an array with a 28-hour rebuild, removing an eight-hour human delay is a meaningful reduction in exposure.
What about RAID 50 or 60?
They stripe across several parity groups, which shortens rebuilds because only one group rebuilds at a time and improves throughput because writes spread across groups. The cost is one or two parity drives per group rather than per array. They are a reasonable middle ground on large drive counts.
Written by
Do The Calculation Team
Do The Calculation
Do The Calculation is built by a small team of data analysts and spreadsheet developers. Where a guide depends on a published formula, standard, or government rule, the calculator it links to names that source directly so you can check the number yourself.
About the team