ZFS and BTRFS

For a long time, the software RAID implementation in the Linux kernel has worked well to protect data against drive failure. It provides great protection against a drive totally failing and against the situation where a drive returns read errors. But what it doesn't offer is protection against silent data corruption (where a disk returns corrupt data and claims it to be good). It also doesn't have good support for the possibility of drive failure during RAID reconstruction.

Drives have been increasing in size significantly, without comparable increases in speed. Modern drives have contiguous read speeds 300 times faster than drives from 1988 but are 40,000 times larger (I'm comparing a recent 4TB SATA disk with a 100M ST-506 disk that can sustain 500K/s reads). So the RAID rebuild time is steadily increasing, while the larger storage probably increases the risk of data corruption.

Currently, there are two filesystems available on Linux that support internal RAID with checksums on all data to prevent silent corruption: ZFS and BTRFS. ZFS is from Sun and has some license issues, so it isn't included in most Linux distributions. It is available from the ZFS On Linux Web site (zfsonlinux.org). BTRFS has no license problems and is included in most recent distributions, but it is at an earlier stage of development. When discussing BTRFS in this article, I concentrate on the theoretical issues of data integrity and not the practical issues of kernel panics (which happen regularly to me but don't lose any data).

For a drive totally to fail (that is, be unable to read any data successfully at all), the most common problem used to be “stiction”. That is when the heads stick to the platters, and the drive motor is unable to spin the disk. This seems to be very uncommon in recent times. I presume that drive manufacturers fixed most of the problems that caused it.

In my experience, the most common reason for a drive to become totally unavailable is due to motherboard problems, cabling or connectors—that is, problems outside the drive. Such problems usually can be fixed but may cause some downtime, and the RAID array needs to keep working with a disk missing.

Serious physical damage (for example, falling on concrete) can cause a drive to become totally unreadable. But, that isn't a problem that generally happens to a running RAID array. Even when I've seen drives fail due to being in an uncooled room in an Australian summer, the result has been many bad sectors, not a drive that's totally unreadable. It seems that most drive “failures” are really a matter of an increasing number of bad sectors.

There aren't a lot of people who can do research on drive failure. An individual can't just buy a statistically significant number of disks and run them in servers for a few years. I couldn't find any research on the incidence of excessive bad sectors vs. total drive failure. It's widely regarded that the best research on the incidence of hard drive “failure” is the Google Research paper “Failure Trends in a Large Disk Drive Population” (research.google.com/pubs/pub32774.html), which although very informative, gives no information on the causes of “failure”. Google defines “failure” as anything other than an upgrade that causes a drive to be replaced. Not only do they not tell us the number of drives that totally died vs. the number that had some bad sectors, but they also don't tell us how many bad sectors would be cause for drive replacement.

Lakshmi N. Bairavasundaram, Garth R. Goodson, Bianca Schroeder, Andrea C. Arpaci-Dusseau and Remzi H. Arpaci-Dusseau from the University of Wisconsin-Madison wrote a paper titled “An Analysis of Data Corruption in the Storage Stack” (research.cs.wisc.edu/adsl/Publications/corruption-fast08.html). That paper gives a lot of information about when drives have corrupt data, but it doesn't provide much information about the case of major failure (tens of thousands of errors), as distinct from cases where there are dozens or hundreds of errors. One thing it does say is that the 80th percentile of latent sector errors per disk with errors is “about 50”, and the 80th percentile of checksum mismatches for disks with errors is “about 100”. So most disks with errors have only a very small number of errors. It's worth noting that this research was performed with data that NetApp obtained by analyzing the operation of its hardware in the field. NetApp has a long history of supporting a large number of disks in many sites with checksums on all stored data.

I think this research indicates that the main risks of data loss are corruption on disk or a small number of read errors, and that total drive failure is an unusual case.

By default, a BTRFS filesystem that is created for a single device that's not an SSD will use “dup” mode for metadata. This means that every metadata block will be written to two parts of the disk. In practice, this can allow for recovering data from drives with many errors. I recently had a 3TB disk develop about 14,000 errors. In spite of such a large number of errors, the duplication of metadata meant that there was little data loss. About 2,000 errors in metadata blocks were corrected with the duplicates, and the 12,000 errors in data blocks (something less than 48M of data) was a small fraction of a 3TB disk. If an older filesystem was used in that situation, a metadata error could corrupt a directory and send all it's entries to lost+found.

ZFS supports even greater redundancy via the copies= option. If you specify copies=2 for a filesystem, then every data block will be written to two different parts of the disk. The number of copies of metadata will be one greater than the number of copies of data, so copies=2 means that there will be three copies of every metadata block. The maximum number of copies for data blocks in ZFS is three, which means that the maximum number of copies of metadata is four.

The paper “An Analysis of Data Corruption in the Storage Stack” shows that for “nearline” disks (that is, anything that will be in a typical PC or laptop), you can expect a 9.5% probability of read errors (latent sector errors) and a 0.466% probability of silent data corruption (checksum mismatches). The typical Linux Journal reader probably can expect to see data loss from hard drive read errors on an annual basis from the PCs owned by their friends and relatives. The probability of silent data corruption is low enough that every user has a less than 50% chance of seeing it on their own PC during their life—unless they purchased one of the disks with a firmware bug that corrupts data.

If you run BTRFS on a system with a single disk (for example, a laptop), you can expect that if the disk develops any errors, they will result in no metadata loss due to duplicate metadata, and any file data that is lost will be reported to the application by a file read error. If you run ZFS on a single disk, you can set copies=2 or copies=3 for the filesystem that contains your most important data (such as /home on a workstation) to decrease significantly the probability that anything less than total disk failure will lose data. This option of providing extra protection for data is a significant benefit for ZFS when compared to BTRFS.

If given a choice between a RAID-1 array with Linux software RAID (or any other RAID implementation that doesn't support checksums) and a single disk using BTRFS, I'd choose the single disk with BTRFS in most cases. That is because on a single disk with BTRFS, the default configuration is to use “dup” for metadata. This means that a small number of disk errors will be unlikely to lose any metadata, and a scrub will tell you which file data has been lost due to errors. Duplicate metadata alone can make the difference between a server failing and continuing to run. It is possible to run with “dup” for data as well, but this isn't a well supported configuration (it requires mixed data and metadata chunks that require you to create a very small filesystem and grow it).

It is possible to run RAID-1 on two partitions on a single disk if you are willing to accept the performance loss. I have a 2TB disk running as a 1TB BTRFS RAID-1, which has about 200 bad sectors and no data loss.

Finally, it's worth noting that a “single disk” from the filesystem perspective can mean a RAID array. There's nothing wrong with running BTRFS or ZFS over a RAID-5 array. The metadata duplication that both those filesystems offer will reduce the damage if a RAID-5 array suffers a read error while replacing a failed disk. A hardware RAID array can offer features that ZFS doesn't offer (such as converting from RAID-1 to RAID-5 and then RAID-6 by adding more disks), and hardware RAID arrays often include a write-back disk cache that can improve performance for RAID-5/6 significantly. There's also nothing stopping you from using BTRFS or ZFS RAID-1 over a pair of hardware RAID-5/6 arrays.

When you replace a disk in Linux Software RAID, the old disk will be marked as faulty first, and all the data will be reconstructed from other disks. This is fine if the other disks are all good, but if the other disks have read errors or corrupt data, you will lose data. What you really need is to have the new disk directly replace the old disk, so the data for the new disk can be read from the old disk or from redundancy in the array, whichever works.

ZFS has a zpool replace command that will rebuild the array from the contents of the old disk and from the other disks in a redundant set. BTRFS supports the same thing with the btrfs replace command. In the most common error situations (where a disk has about 50 bad sectors), this will give you the effect of having an extra redundant disk in the array. So a RAID-5 array in BTRFS or in ZFS (which they call a RAID-Z) should give as much protection as a RAID-6 array in a RAID implementation that requires removing the old disk before adding a new disk. At this time, RAID-5 and RAID-6 support in BTRFS is still fairly new, and I don't expect it to be ready to use seriously by the time this article is published. But the design of RAID-5 in BTRFS is comparable to RAID-Z in ZFS, and they should work equally well when BTRFS RAID-5 code has been adequately tested and debugged.

Hot spare disks are commonly used to allow replacing a disk more quickly than someone can get to the server. The idea is that the RAID array might be reconstructed before anyone even can get to it. But it seems to me that the real benefit of a hot-spare when used with a modern filesystem, such as ZFS or BTRFS, is that the system has the ability to read from the disk with errors as well as the rest of the array while constructing the new disk. If you have a server where every disk bay contains an active disk (which is a very common configuration in my experience), it is unreasonably difficult to support a disk replacement operation that reads from the failing disk (using an eSATA device for the rebuild isn't easy). Note that BTRFS doesn't have automatic hot-spare support yet, but it presumably will get it eventually. In the meantime, a sysadmin has to instruct it to replace the disk manually.

As modern RAID systems (which on Linux servers means ZFS as the only fully functional example at this time) support higher levels of redundancy, one might as well use RAID-Z2 (the ZFS version of RAID-6) instead of RAID-5 with a hot spare, or a RAID-Z3 instead of a RAID-6 with a hot-spare. When a disk is being replaced in a RAID-6/RAID-Z2 array with no hot-spare, you are down to a RAID-5/RAID-Z array, so there's no reason to use a disk as a hot-spare instead of using it for extra redundancy in the array.

The way ZFS works is that the copies= option (and the related metadata duplication) is applied on top of the RAID level that's used for the storage “pool”. So if you use copies=2 on a ZFS filesystem that runs on a RAID-1, there will be two copies of the data on each of the disks. The allocation of the copies is arranged such that it covers different potential failures to the RAID level, so if you had copies=3 for data stored on a three-disk RAID-Z pool, each disk in the pool would have a copy of the data (and parity to help regenerate two other copies). The amount of space required for some of these RAID configurations is impractical for most users. For example, a RAID-Z3 array composed of six 1TB disks would have 3TB of RAID-Z3 capacity. If you then made a ZFS filesystem with copies=3, you would get 1TB of usable capacity out of 6TB of disks. 5/6 disks is more redundancy than most users need.

If data is duplicated in a RAID-1 array, the probability of two disks having errors on matching blocks from independent random errors is going to be very low. The paper from the University of Wisconsin-Madison notes that firmware bugs can increase the probability of corrupt data on matching blocks and suggests using staggered stripes to cover that case. ZFS does stagger some of its data allocation to deal with that problem. Also, it's fairly common for people to buy disks from two different companies for a RAID-1 array to prevent a firmware bug or common manufacturing defect from corrupting data on two identical drives. The probability of both disks in a BTRFS RAID-1 array having enough errors that data is lost is very low. With ZFS, the probability is even lower due to the mandatory duplication of metadata on top of the RAID-1 configuration and the option of duplication of data. At this time, BTRFS doesn't support duplicate metadata on a RAID array.

The probability of hitting a failure case that can't be handled by RAID-Z2 but that can be handled by RAID-Z3 is probably very low. In many deployments, the probability of the server being stolen or the building catching on fire will be greater than the probability of a RAID-Z2 losing data. So it's worth considering when to spend more money on extra disks and when to spend money on better off-site backups.

In 2007, Val Bercovici of NetApp suggested in a StorageMojo interview that “protecting online data only via RAID 5 today verges on professional malpractice” (storagemojo.com/2007/02/26/netapp-weighs-in-on-disks). During the past seven years, drives have become bigger, and the difficulties we face in protecting data have increased. While Val's claim is hyperbolic, it does have a basis in fact. If you have only the RAID-5 protection (a single parity block protecting each stripe), there is a risk of having a second error before the replacement disk is brought on-line. However, if you use RAID-Z (the ZFS equivalent of RAID-5), every metadata block is stored at least twice in addition to the RAID-5 type protection, so if a RAID-Z array entirely loses a disk and then has a read error on one of the other disks, you might lose some data but won't lose metadata. For metadata to be lost on a RAID-Z array, you need to have one disk die entirely and then have matching read errors on two other disks. If disk failures are independent, it's a very unlikely scenario. If, however, the disk failures are not independent, you could have a problem with all disks (and lose no matter what type of RAID you use).

One nice feature of BTRFS and ZFS is the ability to make snapshots of BTRFS subvolumes and ZFS filesystems. It's not difficult to write a cron job that makes a snapshot of your important data every hour or even every few minutes. Then when you accidentally delete an important file, you easily can get it back. Both BTRFS and ZFS can be configured such that files can be restored from snapshots without root access so users can recover their own files without involving the sysadmin.

Snapshots aren't strictly related to the the topic of data integrity, but they solve the case of accidental deletion, which is the main reason for using backups. From a sysadmin perspective, snapshots and RAID are entirely separate issues. From the CEO perspective, “is the system working or not?”, they are part of the same issue.

For a single disk in a default configuration, both BTRFS and ZFS will store two copies of each metadata block. They also use checksums to detect when data is corrupted, which is much better than just providing corrupt data to an application and allowing errors to propagate. ZFS supports storing as many as three copies of data blocks on a single disk, which is a significant benefit.

For a basic RAID-1 installation, BTRFS and ZFS offer similar features by default (storing data on both devices with checksums to cover silent corruption). ZFS offers duplicate metadata as a mandatory feature and the option of duplicate data on top of the RAID configuration.

BTRFS supports RAID-0, which is a good option to have when you are working with data that is backed up well. The combination of the use of BTRFS checksums to avoid data corruption and RAID-0 for performance would be good for a build server or any other system that needs large amounts of temporary file storage for repeatable jobs but for which avoiding data corruption is important.

BTRFS supports dynamically increasing or decreasing the size of the filesystem. Also, the filesystem can be rebalanced to use a different RAID level (for example, migrating between RAID-1 and RAID-5). ZFS, however, has a very rigid way of managing storage. For example, if you have a RAID-1 array in a pool, you can never remove it, and you can grow it only by replacing all the disks with larger ones. Changing between RAID-1 and RAID-Z in ZFS requires a backup/format/restore operation, while on BTRFS, you can just add new disks and rebalance.

ZFS supports different redundancy levels (via the copies= setting) on different “filesystems” within the same “pool” (where a “pool” is group of one or more RAID sets). BTRFS “subvolumes” are equivalent in design to ZFS “filesystems”, but BTRFS doesn't support different RAID parameters for subvolumes at this time.

ZFS supports RAID-Z and RAID-Z2, which are equivalent to BTRFS RAID-5, RAID-6—except that RAID-5 and RAID-6 are new on BTRFS, and many people aren't ready to trust important data to them. There is no feature in BTRFS or planned for the near future that compares with RAID-Z3 on ZFS. There are plans for future development of extreme levels of redundancy in BTRFS at some future time, but it probably won't happen soon.

Generally, it seems that ZFS is designed to offer significantly greater redundancy than BTRFS supports, while BTRFS is designed to be easier to manage for smaller systems.

Currently, BTRFS doesn't give good performance. It lacks read optimization for RAID-1 arrays and doesn't have any built-in support for using SSDs to cache data from hard drives. ZFS has many performance features and is as fast as a filesystem that uses so much redundancy can be.

Finally, BTRFS is a new filesystem, and people are still finding bugs in it—usually not data loss bugs but often bugs that interrupt service. I haven't yet deployed BTRFS on any server where I don't have access to the console, but I have Linux servers running ZFS in another country.