RAID Levels Explained: Usable Capacity and Fault Tolerance

Every RAID level answers the same two questions in a different ratio: how much of the capacity you bought can you actually write to, and how many drives can die before you lose all of it. RAID 5 and RAIDZ1 give you (n-1) times the smallest drive and survive one failure. RAID 6, RAIDZ2 and SHR-2 give you (n-2) and survive two. RAID 10 gives you half. The wrinkle that costs people the most money is the phrase "times the smallest drive": on every layout except Synology SHR, one small drive drags every larger drive down to its size and strands the difference permanently.

What does RAID actually do, and what does it not do?

RAID is a method of spreading data across several drives so that the array keeps serving files when one or more of those drives fails. That is the entire purpose. It is a solution to hardware failure and to nothing else.

The redundancy is bought with parity: mathematical information, computed across the members, from which the contents of a missing drive can be reconstructed. Parity occupies real capacity, always in whole-drive units, which is why a four-drive RAID 6 array holds two drives of data and two drives of protection.

RAID is not a backup. Parity is computed continuously and applied to whatever the filesystem writes, correct or not. Delete a folder and the deletion is protected with perfect redundancy. The same is true of a ransomware run, a bad cable writing corrupted blocks, a fire, a flood and a theft, all of which take every member of the array at once. Read the 3-2-1 backup rule and then come back, because choosing a RAID level before you have a backup plan is optimising the wrong thing.

What RAID buys is uptime and convenience: a drive dies, the array keeps serving, you replace the drive and it rebuilds. Without parity that same failure means restoring everything from backup, which on a 24 TB library is a week of your life even when the backup is perfect.

How much usable capacity does each RAID level give me?

Here is every layout run against the same four 8 TB drives, 32 TB raw. This is the identical drive case, and it is the case where the differences between the families disappear.

Four drives of 8 TB each, 32 TB raw, computed with the same functions that drive our RAID capacity calculator. RAIDZ3 needs five members, so it cannot be built here. "As the OS reports it" is the identical capacity counted in binary tebibytes, which is 9.05 percent smaller by definition. "Rebuild reads" is how much data the array must read from the surviving members to reconstruct one failed drive.

Read the RAID 5, RAIDZ1 and SHR rows next to each other. They are the same three numbers. With identical drives, Synology SHR is exactly RAID 5 and SHR-2 is exactly RAID 6. No vendor magic, no extra terabytes. The SHR band algorithm finds a single band spanning all four drives and applies one drive of parity to it, which is the definition of RAID 5. Anyone telling you SHR is more space-efficient on a matched set of drives is wrong, and the arithmetic above is the proof.

What are RAID 0, RAID 1 and JBOD?

JBOD is not RAID at all. It is a group of drives presented as one volume by filling them in sequence, or simply as separate volumes. Usable capacity is the sum of every drive with nothing lost, and there is no redundancy whatsoever. Its one virtue is that a failure is contained: you lose the files that lived on the dead drive and keep the rest.

RAID 0 stripes data across every drive, splitting each write into chunks written in parallel. Usable capacity is n times the smallest drive, so nothing is lost to parity, and throughput scales with drive count. It also multiplies your risk by drive count: any one of the four drives failing destroys the entire volume, because every file has pieces on every member. Four drives at a 1.5 percent annualised failure rate gives roughly a 6 percent chance of losing the whole array each year. RAID 0 belongs on video scratch volumes and build caches that are fully reproducible, and nowhere else.

RAID 1 writes an identical copy to every member. Usable capacity is one drive regardless of how many you install, and the array survives every drive but one. It is the default in two-bay boxes such as the Synology DS225+, and it has a property the parity levels do not: a rebuild is a straight copy from one surviving drive, so it reads 8 TB rather than 24 TB and finishes in a fraction of the time. That makes RAID 1 unfashionable and quietly excellent.

What is the difference between RAID 5 and RAID 6?

RAID 5 stripes data with one drive's worth of distributed parity and survives one failure. RAID 6 carries a second, independently computed parity and survives two. The second parity costs exactly one drive of capacity, and on the four-drive set above that is the difference between 24 TB and 16 TB usable.

The reason RAID 6 is the default on anything larger than a small array is not the odds of two drives failing on a random Tuesday. It is the rebuild window. Replacing a failed member in a four-drive RAID 5 array forces the array to read every sector of all three survivors, 24 TB in the table above, at a realistic 80 to 150 MB per second while also serving files. On larger drives that becomes days. RAID 5 has zero protection for the entire duration, and it is running the heaviest sustained read job those drives will ever see, on drives the same age as the one that just died.

There is a second, softer argument about unrecoverable read errors. Manufacturers rate NAS drives at one unrecoverable read error per 10^14 or 10^15 bits read, and a large rebuild reads enough bits to make that a live number. Treat that rating carefully: it is a warranty floor rather than a measurement, observed rates across large fleets are considerably better, and both ZFS and mdadm survive a single read error by reporting the affected file rather than aborting the rebuild. Use it to rank layouts, never as a forecast. The full case lives in RAID 5 against RAID 6.

A practical line to hold: single parity up to about four bays with drives no larger than 8 TB, double parity everywhere else. If a rebuild would take more than about a day, you want the second parity drive. Our rebuild time chart tells you which side of that line your drives fall on.

When is RAID 10 worth giving up half the capacity?

RAID 10 is a stripe across mirrored pairs: drives are mirrored two at a time, and the resulting pairs are striped together. Usable capacity is half the drives times the smallest drive, it needs an even number of members, and it survives one failure guaranteed and up to one drive from every pair in the best case. With four drives, two simultaneous failures kill the array one time in three, so the honest number to publish is one.

Half your capacity is a steep price and the payoff is specific. RAID 10 has no parity to compute, so random writes do not carry the read-modify-write penalty that hurts RAID 5 and RAID 6, and its rebuild is a plain copy from the surviving mirror: 8 TB read rather than 16 TB, at full sequential speed, without touching the other pair. For virtual machine images, database volumes and anything with small random writes, that combination is worth the capacity. For a media library, a photo archive or a backup target, it is not, and RAID 6 or RAIDZ2 gives you more space with better tolerance. See RAID 10 against RAID 6.

How is ZFS RAIDZ different from RAID 5 and RAID 6?

The capacity arithmetic is identical: RAIDZ1 is (n-1) times the smallest drive, RAIDZ2 is (n-2), RAIDZ3 is (n-3). What differs is everything around the arithmetic, because ZFS is a filesystem and a volume manager in one rather than a block layer with a filesystem on top.

• Every block is checksummed. ZFS knows whether the data it read back is the data it wrote, and with redundancy present it repairs the bad copy automatically. Classic RAID cannot tell good data from silently corrupted data, because parity only tells it the stripe disagrees, not which member is lying.
• No write hole. RAID 5 and RAID 6 can be left with a stripe half written if power fails mid-write, which is precisely why a correctly sized UPS matters. ZFS writes in complete transactions and has no equivalent state.
• Resilver copies data, not sectors. ZFS reconstructs only the blocks actually in use, so a half-empty pool rebuilds in roughly half the time. mdadm and hardware RAID rebuild every sector whether it holds anything or not.
• Snapshots and send/receive come free with the filesystem, which changes what a backup strategy looks like.

The cost is rigidity. A RAIDZ vdev has historically had a fixed width, so growing a pool has normally meant adding an entire second vdev rather than one drive. RAIDZ expansion, which adds a single drive to an existing vdev, is a recent addition to OpenZFS rather than something that has been there for years, and it redistributes existing data without rewriting the old parity ratio. Plan pool geometry before you buy drives, not after, and see how to migrate to a bigger NAS for what growing actually involves.

What is Synology SHR, and when does it actually help?

SHR (Synology Hybrid RAID) slices the drives into horizontal bands at every distinct drive size and gives each band its own parity group. A band sitting on k drives with p drives of parity yields (k-p) times the band height when k is greater than p, and is stranded entirely when there are not enough drives in that band to protect it. SHR uses one drive of parity, SHR-2 uses two.

Underneath, it is mdadm and Btrfs on standard Linux, not a proprietary format. That matters for recovery: an SHR volume can be assembled and read on an ordinary Linux machine if the Synology unit itself dies.

Here is the same eleven layouts run against 4 + 8 + 12 + 12 TB, 36 TB raw. This is the case that separates SHR from everything else on the list.

Drives of 4, 8, 12 and 12 TB, 36 TB raw. Look at the stranded column: every classic RAID level and every RAIDZ vdev truncates all four members to the 4 TB drive and writes off 20 TB you paid for. SHR strands nothing.

SHR returns 24 TB usable against RAID 5 at 12 TB on the identical hardware, 12 TB more, a gain of 100 percent, with the same single drive of fault tolerance. The bands do it: 0 to 4 TB sits on all four drives and yields 12 TB, 4 to 8 TB sits on three drives and yields 8 TB, 8 to 12 TB sits on two drives and yields 4 TB. Nothing is left over.

Note the SHR-2 row in the same table. Double parity strands 8 TB here, because the top band sits on only two drives and two drives cannot carry two drives of parity. SHR is not a free lunch, it is a band algorithm, and it has edges.

So the rule is simple. Identical drives: SHR gains you nothing and RAID 5 or RAID 6 is the same thing under a different name. Mixed drives, or a plan to upgrade capacity one drive at a time, and SHR is worth real terabytes. Work it through with your own drive set in the RAID capacity calculator, which is the only cross-vendor one: SHR and SHR-2 next to RAIDZ1, RAIDZ2 and RAIDZ3 next to the standard levels, on the same physical drives, including mixed sizes. Then read SHR against RAID 5 for the full comparison.

For completeness, one more model exists and does not fit this table at all: Unraid does not stripe, it stores whole files on individual drives and dedicates one or two entire drives to parity, so drives of any mixed size can be pooled, only the drive being read spins up, and a failure beyond the parity count loses only the files on the drives that died.

How much does the array have to read to rebuild a failed drive?

Rebuild read volume is (n-1) times the smallest drive for single parity and (n-2) times the smallest for double parity. It is the number that decides which layout you should run, and almost nobody publishes it, so here is what it looks like as the array gets wider.

Eight drives of 8 TB each, 64 TB raw. Compare the efficiency column against the four-drive table: RAID 6 improves from 50 percent to 75 percent purely by being wider, which is the strongest argument for one large array over two small ones. Compare the rebuild column against RAID 1 and RAID 10, both of which rebuild by copying a single 8 TB mirror.

Two things fall out of that table. Parity efficiency improves as arrays get wider, so eight drives under RAID 6 keeps 75 percent of the raw capacity while four drives under RAID 6 keeps only 50 percent. And rebuild read volume grows just as fast: a single-parity rebuild across eight 8 TB members reads 56 TB, which at 80 to 150 MB per second is roughly 4 to 8 days. That is why double parity stops being optional somewhere around five or six members, and why the width that makes parity efficient is the same width that makes single parity indefensible.

Is hardware RAID or software RAID better at home?

Hardware RAID puts the parity computation and the array metadata on a dedicated controller card with its own processor and battery-backed cache. Software RAID does the same work on the host CPU, with the metadata written to the drives in a documented, portable format. For a home server, software wins on nearly every axis that matters.

The decisive issue is not performance. A modern CPU computes RAID 6 parity faster than the drives can absorb it, and the ZFS overhead on a four-drive pool is a rounding error on any recent chip. The decisive issue is the controller as its own failure mode. A hardware RAID card writes array metadata in a vendor-specific on-disk format. If the card dies, the drives are intact and unreadable: you generally need the same model of card, sometimes the same firmware revision, to import your own array. That is a single point of failure that parity does not cover and that gets harder to satisfy every year the card is out of production.

Software layouts travel. An mdadm array with Btrfs on top, which is what a Synology SHR volume actually is, can be assembled on any Linux machine with the drives in any order. A ZFS pool imports on any host running ZFS, whatever the hardware. Data portability is a real feature, and it is the answer to the question "what happens when the box dies", which is a question every home server owner eventually gets to ask.

Two related notes. If your motherboard offers RAID in its firmware setup, that is usually "fakeRAID", which is neither: the parity work is done by a driver on the host CPU while the metadata format is still proprietary, so you get the disadvantages of both. And if you buy a hardware controller for the ports rather than the RAID, flash it to plain HBA or IT mode so the card passes drives straight through and lets ZFS or mdadm own them. Deciding between an appliance and a build of your own is covered in NAS or DIY home server.

Which RAID level should I choose?

Match the layout to the shape of the build, not to a reputation.

"n" is the number of drives in the array. Every formula except the two SHR rows truncates each member to the size of the smallest drive. Precalculated figures at every bay count and capacity are in the RAID level comparison chart.

Three shortcuts that hold up in practice. Two bays: RAID 1, because there is no other sensible option and the rebuild is a straight copy. Four bays with drives of 8 TB or smaller: RAID 5, RAIDZ1 or SHR, on the understanding that you have a real backup and that a rebuild will take about a day. Five bays or more, or any array of drives 12 TB and up: double parity, meaning RAID 6, RAIDZ2 or SHR-2, and treat that as a hard rule rather than a preference.

Choose the drives before you choose the layout, because the layout only redistributes what the drives gave you. Four IronWolf Pro 12TB drives under RAID 6 hold more than eight IronWolf 4TB drives under the same layout, in half the bays, drawing less power. Our guide to choosing NAS drives has the cost per terabyte table for that decision, and a four-bay Synology DS925+ or similar appliance will happily run SHR, SHR-2 or classic RAID on whatever you put in it.

Related reading

• RAID capacity calculator for your own drives, including mixed sizes, across every layout at once
• SHR against RAID 5 for exactly where the band algorithm pays
• RAID 5 against RAID 6 for the rebuild exposure argument in full
• TB against TiB chart for why the OS reports 9.05 percent less than the label
• 3-2-1 backup rule because none of these levels is a backup