Understanding ZFS Scrubs and Data Integrity
klarasystems.com68 points by zdw 6 days ago
68 points by zdw 6 days ago
The article is correct but it downplays an important limitation of ZFS scrubs when it talks about how they're different from fsck and chkdsk. As the article says (in different words), ZFS scrubs do not check filesystem objects for correctness and consistency; it only checks that they have the expected checksum and so have not become corrupted due to disk errors or other problems. Unfortunately it's possible for ZFS bugs and issues to give you filesystem objects that have problems, and as it stands today ZFS doesn't have anything that either checks or corrects these. Sometimes you find them through incorrect results; sometimes you discover they exist through ZFS assertion failures triggering kernel panics.
(We run ZFS in production and have not been hit by these issues, at least not that we know about. But I know of some historical ZFS bugs in this area and mysterious issues that AFAIK have never been fully diagnosed.)
A loooong time age (OpenSolaris days) I had a system that had corrupted its zfs. No fsck was available because the developers claimed (maybe still do) that it's unnecessary.
I had to poke around the raw device (with dd and such) to restore the primary superblock with one of the copies (that zfs keeps in different locations on the device). So clearly the zfs devs thought about the possibility of a corrupt superblock, but didn't feel the need to provide a tool to compare the superblocks and restore one from the other copies. That was the point when I stopped trusting zfs.
Such arrogance…
> So clearly the zfs devs thought about the possibility of a corrupt superblock, but didn't feel the need to provide a tool to compare the superblocks and restore one from the other copies.
This mailing list post from 2008 talks about using zdb(8) to mark mark certain uberblocks an invalid so another one would be used:
* https://zfs-discuss.opensolaris.narkive.com/Tx4FaUMv/need-he...
ZDB = ZFS debugger. It's been there since the original Solaris release of ZFS.
> That was the point when I stopped trusting zfs.
As opposed to trusting other file systems and volume managers, which do not have checksums, and so you wouldn't even know about the problem in the first place?
That's a fine fit of pique - and I once had an awkward file on one of my zfs pools, about three pools ago - but how does it leave you better off, if you want what zfs offers?
"Scrubs differ significantly from traditional filesystem checks. Tools such as fsck or chkdsk examine logical structures and attempt to repair inconsistencies related to directory trees, allocation maps, reference counts, and other metadata relationships. ZFS does not need to perform these operations during normal scrubs because its transactional design ensures metadata consistency. Every transaction group moves the filesystem from one valid state to another. The scrub verifies the correctness of the data and metadata at the block level, not logical relationships."
A scrub literally reads the object from disk. And, for each block, the checksums are read up the tree. The object is therefore guaranteed to be correct and consistent at least re: the tree of blocks written.
> Unfortunately it's possible for ZFS bugs and issues to give you filesystem objects that have problems
Can you give a more concrete example of what you mean? It sounds like you have some experience with ZFS, but "ZFS doesn't have an fsck" is also some truly ancient FUD, so you will forgive my skepticism.
I'm willing to believe that you request an object and ZFS cannot return that object because of ... a checksum error or a read error in a single disk configuration, but what I have never seen is a scrub that indicates everything is fine, and then reads which don't return an object (because scrubs are just reads themselves?).
Now, are things like pool metadata corruption possible in ZFS? Yes, certainly. I'm just not sure fsck would or could help you out of the same jam if you were using XFS or ext4. AFAIK fsck may repair inconsistencies but I'm not sure it can repair metadata any better than ZFS can?
Two examples that I can find are https://github.com/openzfs/zfs/issues/7910, where very old versions of ZFS appear to have quietly written slightly incorrect ACL information, and https://bugs.launchpad.net/ubuntu/+source/zfs-linux/+bug/190... where Ubuntu 21.10 shipped with a bug that created corrupted ZFS filesystems. I believe https://www.illumos.org/issues/9847 may be another example of this, although less severe, where ZFS leaked disk space under some circumstances.
> Can you give a more concrete example of what you mean?
There's been several instances. For example, the send/receive code has had bugs leading to cases[1] where the checksum and hence scrub look fine but the data is not.
edit: the recent block cloning has also had some issues, eg[2][3].
I'm pretty sure it's also possible for hardware errors like bad memory to cause the data to get corrupted but the checksum gets computed on the corrupted data, thus it looks ok when scrubbed.
[1]: https://github.com/openzfs/zfs/issues/4809
Generally, it's possible to have data which is not corrupted but which is logically inconsistent (incorrect).
Imagine that a directory ZAP has an entry that points to a bogus object ID. That would be an example. The ZAP block is intact but its content is inconsistent.
Such things can only happen through a logical bug in ZFS itself, not through some external force. But bugs do happen.
If your search through OpenZFS bugs you will find multiple instances. Things like leaking objects or space, etc. That's why zdb now has support for some consistency checking (bit not for repairs).
> Imagine that a directory ZAP has an entry that points to a bogus object ID. That would be an example. The ZAP block is intact but its content is inconsistent.
The above is interesting and fair enough, but a few points:
First, I'm not sure that makes what seems to be the parent's point -- that scrub is an inadequate replacement for an fsck.
Second, I'm really unsure if your case is the situation the parent is referring to. Parent seems to be indicating actual data loss is occurring. Not leaking objects or space or bogus object IDs. Parent seems to be saying she/he scrubs with no errors and then when she/he tries to read back a file, oops, ZFS can't.
Imagine a race condition that writes a file node where a directory node should be. You have a valid object with a valid checksum, but it's hooked into the wrong place in your data structure.
> Imagine a race condition that writes a file node where a directory node should be. You have a valid object with a valid checksum, but it's hooked into the wrong place in your data structure.
A few things: 1) Is this an actual ZFS issue you encountered or is this a hypothetical? 2) And -- you don't imagine this would be discovered during a scrub? Why not? 3) But -- you do imagine it would be discovered and repaired by an fsck instead? Why so? 4) If so, wouldn't this just be a bug, like a fsck, not some fundamental limitation of the system?
FWIW I've never seen anything like this. I have seen Linux plus a flaky ALPM implementation drop reads and writes. I have seen ZFS notice at the very same moment when the power dropped via errors in `zpool status`. I do wonder if ext4's fsck or XFS's fsck does the same when someone who didn't know any better (like me!) sets the power management policy to "min_power" or "med_power_with_dipm".
>HDDs typically have a BER (Bit Error Rate) of 1 in 1015, meaning some incorrect data can be expected around every 100 TiB read. That used to be a lot, but now that is only 3 or 4 full drive reads on modern large-scale drives. Silent corruption is one of those problems you only notice after it has already done damage.
While the advice is sound, this number isn't the right number for this argument.
That 10^15 number is for UREs, which aren't going to cause silent data corruption -- simple naive RAID style mirroring/parity will easily recover from a known error of this sort without any filesystem layer checksumming. The rates for silent errors, where the disk returns the wrong data that benefit from checksumming, are a couple of orders of magnitude lower.
This stat is also complete bullshit. If it were true, your scrubs of any 20+TB pool would get at least corrected errors quite frequently. But this is not the case.
The consumer grade drives are often given an even lower spec of 1 in 1e14. For a 20TB drive, that's more than one error every scrub, which does not happen. I don't know about you, but I would not consider a drive to be functional at all if reading it out in full would produce more than one error on average. Pretty much nothing said on that datasheet reflects reality.
This is pure theory. Ber shouldn't be counted per sector etc? We shouldn't tread all disk space as single entity, IMO
Why would that make a difference unless some sectors have higher/lower error rates than others?
> HDDs typically have a BER (Bit Error Rate) of 1 in 1015, meaning some incorrect data can be expected around every 100 TiB read. That used to be a lot, but now that is only 3 or 4 full drive reads on modern large-scale drives
I remember this argument way back 16 years ago when the "Why RAID 5 stops working in 2009" article[0] blew up. It's BS. Those aren't the actual average error rates. Those are inflated error rates below which the manufacturer does not want to bother supplying a warranty for.
I have a pool with 260 TB worth of 10/14 TB disks in it 80% full, with monthly scrubs going back years. Not a single checksum error, and in total something like 30 reallocated sectors seen in SMART (half of those on a 7 year old drive).
[0] https://www.zdnet.com/article/why-raid-5-stops-working-in-20...
>Most systems that include ZFS schedule scrubs once per month. This frequency is appropriate for many environments, but high churn systems may require more frequent scrubs.
Is there a more specific 'rule of thumb' for scrub frequency? What variables should one consider?
I scrub once a quarter because scrubs take 11 days to complete. I have 8x 18TB raidz2 pool, and I keep a couple of spare drives on hand so I can start a resilver as soon as an issue crops up.
In the past, I've gone for a few years between scrubs. One system had a marginal I/O setup and was unreliable for high streaming load. When copying the pool off of it, I had to throttle the I/O to keep it reliable. No data loss though.
Scrubs are intensive. They will IMO provoke failure in drives sooner than not doing them. But they're the kind of failures you want to bring forward if you can afford the replacements (and often the drives are under warranty anyway).
If you don't scrub, eventually you generally start seeing one of two things: delays in reads and writes because drive error recovery is reading and rereading to recover data; or, if you have that disk behaviour disabled via firmware flags (and you should, unless you're reslivering and on your last disk of redundancy), you see zfs kicking a drive out of the pool during normal operations.
If I start seeing unrecoverable errors, or a drive dropping out of the pool, I'll disable scrubs if I don't have a spare drive on hand to start mirroring straight away. But it's better to have the spares. At least two, because often a second drive shows weakness during resilver.
There is a specific failure mode that scrubs defend against: silent disk corruption that only shows up when you read a file, but for files you almost never read. This is a pretty rare occurrence - it's never happened to me in about 50 drives worth of pools over 15 years or so. The way I think about this is, how is it actionable? If it's not a failing disk, you need to check your backups. And thus your scrub interval should be tied to your backup retention.
Once a month might be too high because HDDs are rated at ~ 180 TB workload/year. Remember, the workload/year limit includes read & writes and doesn't vary much by capacity, so a 10 TB HDD scrubbed monthly consumes 67% of the workload, let alone any other usage.
Scrubbing every quarter is usually sufficient without putting high wear on the HDD.
A scrub only reads allocated space, so in your 10TB example, a scrub would only read whatever portion of that 10TB is actually occupied by data. It's also usually recommended to keep your usage below 80% of the total pool size to avoid performance issues, so the worst case in your scenario would be more like ~53% assuming you follow the 80% rule.
Is the 80% rule real or just passed down across decades like other “x% free” rules? Those waste enormous amounts of resources on modern systems and I kind of doubt ZFS actually needs a dozen terabytes or more of free space in order to not shit the bed. Just like Linux doesn’t actually need >100 GB of free memory to work properly.
In practice you see noticeable degradation of performance for streaming reads of large files written after 85% or so. Files you used to be able to expect to get 500+MB/sec could be down to 50MB/sec. It's fragmentation, and it's fairly scale invariant, in my experience.
Speaking strictly about ZFS internal operations, the free space requirement is closer to 5% on current ZFS versions. That allows for CoW and block reallocations in real-world pools. Heavy churn and very large files will increase that margin.
Once a month seems like a reasonable rule of thumb.
But you're balancing the cost of the scrub vs the benefit of learning about a problem as soon as possible.
A scrub does a lot of I/O and a fair amount of computing. The scrub load competes with your application load and depending on the size of your disk(s) and their read bandwidth, it may take quite some time to do the scrub. There's even maybe some potential that the read load could push a weak drive over the edge to failure.
On my personal servers, application load is nearly meaningless, so I do an about monthly scrub from cron which I think will only scrub one zpool at a time per machine, which seems reasonable enough to me. I run relatively large spinning disks, so if I scrubbed on a daily basis, the drives would spend most of the day scrubbing and that doesn't seem reasonable. I haven't run ZFS in a work environment... I'd have to really consider how the read load impacted the production load and if scrubbing with limits to reduce production impact would complete in a reasonable amount of time... I've run some systems that are essentially alwayd busy and if a scrub would take several months, I'd probably only scrub when other systems indicate a problem and I can take the machine out of rotation to examine it.
If I had very high reliability needs or a long time to get replacement drives, I might scrub more often?
If I was worried about power consumption, I might scrub less often (and also let my servers and drives go into standby). The article's recommendation to scan at least once every 4 months seems pretty reasonable, although if you have seriously offline disks, maybe once a year is more approachable. I don't think I'd push beyond that, lots of things don't like to sit for a year and then turn on correctly.
Once a month is fine ("/etc/cron.monthly/zfs-scrub"):
#!/bin/bash
#
# ZFS scrub script for monthly maintenance
# Place in /etc/cron.monthly/zfs-scrub
POOL="storage"
TAG="zfs-scrub"
# Log start
logger -t "$TAG" -p user.notice "Starting ZFS scrub on pool: $POOL"
# Run the scrub
if /sbin/zpool scrub "$POOL"; then
logger -t "$TAG" -p user.notice "ZFS scrub initiated successfully on pool: $POOL"
else
logger -t "$TAG" -p user.err "Failed to start ZFS scrub on pool: $POOL"
exit 1
fi
exit 0Didn't know about the logger script, looks nice. Can it wrap the launch of the scrub itself so that it logs like logger too, or do you separately track its stdout/stderr when something happens?
update: figured how you can improve that call to add logs to logger
Scrub doesn't log anything by default, you run it and it returns quickly... you have to get the results out of zpool status or through zed.
That script might do with the "-w" parameter passed to scrub. Then "zpool scrub" won't return until the scrub is finished.
The cost of a scrub is just a flurry of disk reads and a reduction in performance during a scrub.
If this cost is affordable on a daily basis, then do a scrub daily. If it's only affordable less often, then do it less often.
(Whatever the case: It's not like a scrub causes any harm to the hardware or the data. It can run as frequently as you elect to tolerate.)
With HDDs, it's also mechanical wear and increased chance of a failure. SSDs are not fully immune to increased load either.
Is there any evidence that suggests that reading from a hard drive (instead of it just spinning idle) increases physical wear in any meaningful way? Likewise, is there any evidence of this for solid-state storage?
Yes. Hard drives have published "Annualized Workload Rate" ratings, which are in TB/year, and the manufacturers state there is no difference between reads and writes for the purpose of this rating.
(https://www.toshiba-storage.com/trends-technology/mttf-what-...)
For SSDs, writes matter a lot more. Reads may increase the temperature of the drive, so they'll have some effect, but I don't think I've seen a read endurance rating for an SSD.
Total pool size and speed. Less data scrubs faster, as do faster disks or disk topology (a 3 way stripe of nvme will scrub faster than a single sata ssd)
For what its worth, I scrub daily mostly because I can. It's completely overkill, but if it only takes half an hour, then it can run in the middle of the night while I'm sleeping.