How long does a RAID 6 or RAIDZ2 rebuild take with 20 TB drives?

Why is the drive's top speed only a best case?

Because the top speed holds on the outer tracks only. Western Digital's HC560 data sheet places its 291 MB/s maximum about 10% into the drive. Inner tracks pass under the head more slowly, so a full pass averages less.

The replacement drive is the bottleneck. Every rebuilt block ends up on that one drive, so the array cannot rebuild faster than it writes. Ten drives reading in parallel do not change that.

A rebuild can never beat one drive writing from end to end.

Why does an mdadm rebuild run slower than the drive?

Because md caps it. In the Linux kernel source, speed_limit_max defaults to 200,000 KiB/s, about 204.8 MB/s, and speed_limit_min to 1,000 KiB/s. The md manual explains the pair: md slows recovery when other activity is happening, and runs at full speed when the array is idle.

So a busy array can rebuild near the floor. At 1,000 KiB/s, a 20 TB drive would take about 226 days. Real arrays sit between the two limits, set by how much backup or read traffic competes.

You can watch and steer it:

# progress, speed and time left
cat /proc/mdstat

# current limits, in KiB/s
sysctl dev.raid.speed_limit_min dev.raid.speed_limit_max

# let the rebuild use more of the disks during a quiet period
sysctl -w dev.raid.speed_limit_min=100000
sysctl -w dev.raid.speed_limit_max=500000

Is a ZFS resilver faster than a RAID 6 rebuild?

Often, when the pool is not full. The OpenZFS manual says resilvering only examines data that ZFS knows to be out of date. A pool filled to half rebuilds about half the bytes that mdadm would copy.

The reads are less tidy, though. The OpenZFS dRAID guide notes that raidz must use a healing resilver, which walks the block tree with random reads. A full RAIDZ2 pool can therefore take longer than the bytes suggest.

dRAID changes the picture for large arrays. Its sequential rebuild to a distributed spare scales with the number of disks divided by the stripe width. A scrub then checks the checksums. With ten drives, plain RAIDZ2 is the usual choice; the trade-offs are on ZFS or hardware RAID.

Watch a resilver with zpool status. The zfs_resilver_min_time_ms tunable, 1,500 ms by default in the current manual, sets how long ZFS works on it between transaction groups.

How risky is the rebuild window?

Small for RAID 6, larger for RAID 5. During a RAID 6 rebuild the array still has one parity left, so one more failure is survivable.

Backblaze measured a 1.41% lifetime annualized failure rate across its drives in its Q2 2026 report. Over a 27-hour rebuild, nine remaining drives give about a 0.04% chance of another failure. Over 111 hours, about 0.16%.

Unreadable sectors weigh more. Seagate rates the Exos X20 at one per 10^15 bits read, at most. Rebuilding from nine 20 TB drives reads 1.44 × 10^15 bits, so the spec allows more than one bad sector. RAID 6 repairs it from the second parity; RAID 5 has nothing left. The full comparison is on RAID 6 or RAID 10.

Of 23 storage-server sellers Basalthold checked on 2 November 2026, 8 listed a plan of 100 TB or more. At that size, each drive rebuild takes most of a day even at full speed.

How do I shorten a rebuild on a rented server?

Start it early and give it the disks.

• Replace the drive fast. When a drive fails, open a ticket: the team has it replaced at no cost, and rebuilding the array on the new drive is up to your setup. The clock starts when the new drive is in.
• Pause heavy jobs. Move large backup runs out of the rebuild window, or raise speed_limit_min at night.
• Keep the pool below full. ZFS rebuilds only the data stored, so free space shortens the resilver.
• Keep a second copy. A rebuild protects against one more failure, not against a fire or a mistake: A storage server is not a backup of itself: we keep no copy of your data anywhere else, so a second copy in another place stays your job. A target for it is on offsite backup for a NAS.