Enterprise Techie
Subscribe

RAID 5 vs RAID 6: Parity, Rebuilds and How to Choose

David Davis
RAID 5 vs RAID 6: Parity, Rebuilds and How to Choose

RAID 5 spreads data and a single parity block across every drive in the group and survives one drive failure; RAID 6 adds a second, independent parity block to each stripe and survives any two. RAID 5 needs at least three drives and gives the usable capacity of all but one; RAID 6 needs at least four and gives all but two. The gap matters most during a rebuild: a degraded RAID 5 group has no redundancy left until the replacement drive is rebuilt, while RAID 6 can still lose another drive.

More storage analysis is listed in Enterprise Techie’s topic index.

In this guide

RAID 5 vs RAID 6 vs RAID 10 vs RAID 60 at a glance

RAID 5RAID 6RAID 10RAID 60
Minimum drives3448
Usable capacity (n equal drives)n minus 1n minus 2Half of n (two-way mirrors)n minus 2 per RAID 6 group
Drive failures toleratedAny 1Any 21 per mirrored pair2 per RAID 6 group
Disk I/Os for a one-strip write (IBM)4: read and write data and parity6: two parity blocks2: both mirror copiesAs RAID 6, within one group
Typical use in vendor guidanceHigh read, low write ratesHigh read, low small-block write rates; archive and backupDatabases, performance-sensitive workloadsLarge volumes, including over 32 drives

Minimums come from Dell’s RAID reference and Broadcom’s MegaRAID user guide. Software RAID can differ: OpenZFS allows a single-parity group of two devices.

How RAID 5 and RAID 6 parity work

The SNIA dictionary defines RAID 5 as storing “stripes of ‘n’ logical blocks of data and one logical block of parity across a set of ‘n+1’ independent storage devices,” with parity interleaved across the devices. Parity is the exclusive-or (XOR) of the stripe’s data blocks, so a failed drive’s blocks can be recalculated from parity and the surviving data. Per the Linux md manual, rotating parity across drives, unlike RAID 4’s single parity drive, leaves “less contention” between writes.

RAID 6 keeps two independent parity blocks per stripe, which Broadcom labels P and Q. SNIA’s definition only requires surviving “any two storage device failures without data loss,” so the math varies; Microsoft’s dual parity uses Reed-Solomon codes.

Both share the write hole: an interrupted write leaves data and parity inconsistent, causing “a false regeneration” after a later drive failure. Linux md offers a Partial Parity Log or journal device against it, and OpenZFS says RAID-Z does not suffer from it.

Usable capacity, with a worked example

Broadcom states that RAID 5 usable capacity equals “the number of drives in the array, minus one, into the capacity of the smallest drive in the array,” and RAID 6 the same “minus two,” so larger drives mixed into a group waste the difference. For eight 20 TB drives:

  • RAID 5: 7 x 20 TB = 140 TB usable.
  • RAID 6: 6 x 20 TB = 120 TB usable.
  • RAID 10: four mirrored pairs, 4 x 20 TB = 80 TB usable.
  • RAID 60: two four-drive RAID 6 groups, 2 x 2 x 20 TB = 80 TB usable.

These are nominal figures. Seagate notes that a TB in drive capacity is one trillion bytes and that an operating system “may use a different standard of measurement and report a lower capacity.”

The write penalty: why parity costs extra I/O

Changing part of a stripe means updating its parity too. The kernel’s md documentation says that for non-full-stripe writes, “MD must read old data before the new parity can be calculated,” and those reads “hurt write throughput.” This read-modify-write cycle is the write penalty.

IBM’s Redpaper on RAID-6 for its DS5000 arrays counts the steps for a write that changes one strip: read the old data, read the old parity, write the new parity, write the new data, “a total of four I/O operations to the HDDs.” RAID 6 reads and updates both parity blocks, “six operations in total.” RAID 10 writes the changed data to both drives of a mirrored pair.

Full-stripe writes skip the reads, which is why md’s write-back journal holds data until it fills a stripe. Broadcom calls RAID 5 “Not well suited to tasks requiring lots of small writes or small block write operations.”

Degraded mode, rebuilds and the exposure window

After a drive fails, the array runs in degraded mode, still serving reads and writes and calculating the failed drive’s data as needed. A rebuild is “the regeneration and writing onto one or more replacement devices of all of the data that was on a failed device,” and regeneration uses “check data and user data from surviving members.” Rebuilding one drive means reading from every remaining drive in the group.

Until it finishes, RAID 5 has no redundancy. Seagate’s RAID Manager documentation says that if a second drive fails before synchronization completes, “all data in the array is lost.” Broadcom reports a RAID 6 group missing one drive as “Partially Degraded” and says RAID 6 “Can survive the loss of two drives or the loss of a drive while another drive is being rebuilt.” A hot spare shortens the window, since the rebuild “can start automatically when a drive fails.”

Why larger drives mean longer rebuilds

A rebuild regenerates everything the failed drive held, so twice the data means twice the blocks to recompute and write. Rebuilds also share drives with production I/O: md slows recovery “if other activity is happening,” and Broadcom’s default rebuild rate is 30 percent of compute cycles. NetApp’s ONTAP documentation says “Local tiers that contain larger disks have a higher possibility of concurrent disk failures” and makes triple-parity RAID-TEC the default for capacity HDD tiers with disks of 6 TB or larger.

Unreadable sectors during a rebuild

A working drive can still fail to return a sector. Manufacturers state this per bits read: Seagate’s Exos data sheet (DS2046.1-2512, © 2025) lists “Nonrecoverable Read Errors per Bits Read” as “1 sector per 10E15” for the 32 TB Exos ST32000NM004K. That is a specification, not a forecast for any rebuild.

In a healthy array, the bad sector is rebuilt from redundancy: SNIA says regeneration “may be used to recover data from an unrecoverable media error.” A degraded RAID 5 group has spent that redundancy, so an unreadable sector on a surviving drive leaves its stripe two pieces short with one parity block. RAID 6 still has a second. Since a rebuild reads every surviving drive, the bits read grow with drive capacity.

Scrubbing finds these errors first. The md manual calls it “valuable to regularly read all blocks on all devices in an array so as to catch such bad blocks early”; MegaRAID controllers run patrol reads.

RAID 10 and RAID 60 compared

RAID 10 stripes across mirrored pairs and survives “a single storage device failure in each RAID 1 set.” Writes involve no parity; md rebuilds it by copying surviving originals. Dell notes it survives multiple failures only “as long as two drives of the same mirrored pair do not fail.”

RAID 60 stripes across two or more RAID 6 groups. Broadcom says it “can survive the loss of two drives in each of the RAID 6 sets without losing data” and is “Also used when a virtual drive of greater than 32 drives is needed.”

RAID 6 vs RAID 10 trades capacity against write cost. With eight drives, RAID 6 yields six drives of capacity and survives any two failures; RAID 10 yields four and survives two only if they hit different pairs.

Modern alternatives: RAID-Z, dRAID, Storage Spaces and erasure coding

  • OpenZFS RAID-Z1, Z2, Z3. RAIDZ is “a variation on RAID-5” with single, double or triple parity. N disks of size X with P parity hold “approximately (N-P)*X bytes”; the docs recommend 3 to 9 disks per group, and a group can be widened by attaching a disk.
  • OpenZFS dRAID. Added in OpenZFS 2.1.0, dRAID builds in distributed hot spares “which allows for faster resilvering,” spreading rebuild I/O across all surviving drives. Its fixed stripe width costs capacity with small blocks.
  • Microsoft Storage Spaces. A standalone parity space needs three disks and suits workloads “that are highly sequential, such as archive or backup.” In Storage Spaces Direct, single parity “most closely resembles RAID-5” and is discouraged; dual parity “most closely resembles RAID-6.”
  • NetApp ONTAP. RAID-DP handles two simultaneous disk failures per RAID group and RAID-TEC three, with minimums of 5 and 7 disks.
  • Distributed erasure coding. SNIA says erasure coding suits “large-scale, distributed storage systems that require a configurable balance between storage efficiency, performance, and rebuild speed.” See what is HCI for parity across nodes and object vs block storage for object stores.

RAID is not a backup

Dell is explicit: “The RAID technology is not a backup solution. It does not replace a good data backup solution for data retention and security.” RAID protects against drive failure; a deleted file, corrupted write or ransomware encryption is stored like any other change. Keep separate copies, run restore tests and isolate one copy, as in air-gapped backup.

How to choose by workload and drive size

  • Small random writes (write-heavy databases, busy VM datastores): RAID 10. Broadcom calls RAID 6 “Not well suited to tasks requiring numerous small and/or random write operations,” and Microsoft recommends mirroring “for most performance-sensitive workloads.”
  • Read-heavy or sequential capacity (file shares, archive, backup repositories): RAID 6, which Broadcom suggests for “high read request rates but low random or small block write rates.”
  • Large-capacity HDDs: dual parity at minimum, and triple parity (RAID-Z3, dRAID3, RAID-TEC) for the largest. Keep RAID 5 to small groups of smaller drives holding data that is backed up or easy to recreate.
  • Many drives: several RAID 6 groups (RAID 60) or raidz vdevs, not one wide group. NetApp says larger RAID groups increase reconstruction time and “the probability of a multiple disk failure within the same RAID group.”
  • Always: a hot spare, scheduled scrubs, drive-failure alerts and a tested backup.

Frequently asked questions

Is RAID 6 better than RAID 5?

For protection, yes. RAID 6 survives any two drive failures, or a failed drive plus an unreadable sector during the rebuild; RAID 5 survives one. RAID 6 costs one more drive of capacity and an extra parity update on every write.

How many drives can fail in RAID 6?

Any two drives in the group, per SNIA and the Linux md manual. Seagate notes that a third failure before the rebuild completes loses all data in the array. In RAID 60, the limit is two per RAID 6 group.

Is RAID 5 safe?

It survives one drive failure, then has no redundancy until the rebuild finishes. A second failure or an unreadable sector during that window can lose data, and larger drives lengthen the window. It suits small groups with tested backups better than large-capacity HDDs.

RAID 6 vs RAID 10: which should you use?

RAID 10 for write-heavy, performance-sensitive workloads, at half the raw capacity. RAID 6 for capacity and read-heavy data, with any two drive failures covered.

Sources

Last checked: October 2026. Guidance and product features change; check each source before you rely on it.