Hmm.. while I do appreciate the higher capacity, the 8TB drives coming in at half the price seem to be the much better deal. At this point I could reduce my 5x2TB RAID5 into a 2x8TB RAID1 and gain quite a bit of confidence.
2x8TB RAID1 would be a performance sacrifice but you would gain A LOT of much needed confidence as that 5x2TB RAID5 is one drive failure and a near guaranteed URE waiting to happen.
Near guaranteed URE? That's complete nonsense. I've done a bunch of testing myself and find nothing of this sort to be remotely true.
I built a 10x2TB MD RAID 5 array, filled it with data and then kept removing a drive and rebuilding the array, then verifying all the data. I rebuilt it and verified it 20 times before I needed the disks for something else. I never came across a URE.
I also have a 12x4TB ZFS pool that I scrub twice a month. ZFS notified you of UREs and repairs them when they are found. I've scrubbed my 70% filled pool more than 100 times with no URE encountered.
URE's are not as common as you apparently think they are.
"URE's are not as common as you apparently think they are."
More to the point they're nowhere near as common as the specsheets imply they are. The specsheet numbers (unchanged for a decade or two despite all the capacity increases) imply a failure rate of once per ~12.5TB (100 terrabits); which would make RAIDs above a few terabytes a crapshoot to rebuild and ones a few times larger nearly impossible. The reddit post I've linked has some actual test run data from someone else whose results were the same as yours. Some of the replies, eg the one claiming that the average is dominated by much less frequent failure modes that generate large numbers of errors at once, are interesting as well. Either way the naive prediction of how big an array you can rebuild before being likely doomed by a URE is much too small. Rebuilding an array can fail today just like it could a decade or two ago; but the apocalyptic predictions from early this century never came to pass.
Which specsheets show 1 URE in 12.5TB. All the specsheets for my disks say < 1 in 10^15 as in *less than*. So If i see 1 URE in 100TB or 1 URE in 1PB, that's certainly *less than* 1 in 10^15.
The implication is there because there exist drives with the <1 in 10^15 (1 error per 1000TB) spec. Since these drives are not expected to be that reliable, the implication is that there is a reasonable chance of an error somewhere between 100TB and 1000TB for a drive rated <1 in 10^14.
Of course in reality, other factors come into play. Personally, I chose RAID6 for my NAS. Spinning storage is cheap, so why not.
I think the error is a percent chance, which doesn't accumulate the same way as a "1 per xxTB written" would accumulate.
It's probably something like (1-10^-15)^(10^15) which is about 37%, meaning that if you wrote about 113TBs (10^15 is about 113TB, unless you're one of THOSE people in which case it's 125TB) then you can expect about 1 URE. If you did that but didn't get a URE, and wrote another 113TBs, then expect the odds of not getting a URE to go from 37% to 13.5%.
ZFS has issues of its own. For example, they require ECC system RAM for reliability to a far greater extent than most file systems. Performance can be much lower (because of ZFS's design, not because of hardware RAID performance). Operating system support is somewhat more limited. Of course, ZFS and hardware RAID are not mutually exclusive or even necessarily competitors.
That sounds stupid, the creator of ZFS pointed out that you don't have to use ECC for reliability, ECC is just "nice to have" not a requirement. You just wanted to say a bad thing about ZFS, I wonder what OS you use?
URE's aren't the main problem with RAID5. Controller faults are the main problem with RAID5. Go to any server or enterprise focused forum and note the hostility towards RAID5, and it's well deserved.
RAID5/6 essentially encrypts your data in the form of a parity stripe with very poor data correction unless you have a really high end controller. If the controller gets too warm, or has other issues the parity stripe can get corrupted causing a cascade of data integrity issues. Since RAID 6 creates two logical parity writes this acts as a form of fault tolerance which is why RAID6 doesn't get the ire of RAID5.
RAID 5 only had logical justification back in the 90's when a 9GB drive cost $700. Otherwise, RAID5 failures on servers is still my #1 form of data corruption. If you have't worked on enough storage to see the issue it's not my problem.
If you care about your data you need a backup anyway. Then the chance of both your primary storage and your backup failing at the same time is pretty low even with 12 disks RAID.
The impression that an unrecoverable error will necessarily be corrected by any RAID controller is not necessarily an accurate one. In the event of an individual disk error, RAID1 may be at greater risk because the controller has no way of knowing which copy is correct. RAID5/6 can derive a data mismatch through the parity information on other drives (not that all controllers necessarily do this). One great thing about ZFS is that it stores checksums on parent blocks, allowing detection of data corruption regardless of the source.
It is interesting to note that WD has improved the power consumption of the 10TB drives over the older 8TB drives. We are asking how exactly WD is doing that, as details were not given with the press release.
The Helium makes all the platters easier to spin (ie use less power).
"The WD Pro 10 TB drive is engineered for personal or small business NAS systems with up to eight bays, is optimized for mixed workloads and has a 5400 RPM spindle speed. By contrast, the WD Red Pro 10 TB is aimed at medium business and enterprise-class NAS systems up to 16 bays...."
Is this a typo? Shouldn't it be "The WD Red 10TB is engineered for personal...."?
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
Facebook
Twitter
RSS
19 Comments
Back to Article
bill.rookard - Monday, May 22, 2017 - link
Hmm.. while I do appreciate the higher capacity, the 8TB drives coming in at half the price seem to be the much better deal. At this point I could reduce my 5x2TB RAID5 into a 2x8TB RAID1 and gain quite a bit of confidence.Space Jam - Monday, May 22, 2017 - link
5x2TB RAID5 is a pretty brave decision.2x8TB RAID1 would be a performance sacrifice but you would gain A LOT of much needed confidence as that 5x2TB RAID5 is one drive failure and a near guaranteed URE waiting to happen.
SirMaster - Monday, May 22, 2017 - link
Near guaranteed URE? That's complete nonsense. I've done a bunch of testing myself and find nothing of this sort to be remotely true.I built a 10x2TB MD RAID 5 array, filled it with data and then kept removing a drive and rebuilding the array, then verifying all the data. I rebuilt it and verified it 20 times before I needed the disks for something else. I never came across a URE.
I also have a 12x4TB ZFS pool that I scrub twice a month. ZFS notified you of UREs and repairs them when they are found. I've scrubbed my 70% filled pool more than 100 times with no URE encountered.
URE's are not as common as you apparently think they are.
DanNeely - Monday, May 22, 2017 - link
"URE's are not as common as you apparently think they are."More to the point they're nowhere near as common as the specsheets imply they are. The specsheet numbers (unchanged for a decade or two despite all the capacity increases) imply a failure rate of once per ~12.5TB (100 terrabits); which would make RAIDs above a few terabytes a crapshoot to rebuild and ones a few times larger nearly impossible. The reddit post I've linked has some actual test run data from someone else whose results were the same as yours. Some of the replies, eg the one claiming that the average is dominated by much less frequent failure modes that generate large numbers of errors at once, are interesting as well. Either way the naive prediction of how big an array you can rebuild before being likely doomed by a URE is much too small. Rebuilding an array can fail today just like it could a decade or two ago; but the apocalyptic predictions from early this century never came to pass.
https://www.reddit.com/r/zfs/comments/3gpkm9/stati...
SirMaster - Monday, May 22, 2017 - link
Which specsheets show 1 URE in 12.5TB. All the specsheets for my disks say < 1 in 10^15 as in *less than*. So If i see 1 URE in 100TB or 1 URE in 1PB, that's certainly *less than* 1 in 10^15.SirMaster - Monday, May 22, 2017 - link
I miss-wrote on my previous comment.All the datasheets that I have seen say < 1 in 10^14.
See WD Red for datasheet for these 10TB disks:
https://www.wdc.com/content/dam/wdc/website/downlo...
It does not say equal to 1 in 10^14, so there is no reason to think thing you are anywhere near guaranteed to get a URE in 12.5TB.
Maltz - Monday, May 22, 2017 - link
The implication is there because there exist drives with the <1 in 10^15 (1 error per 1000TB) spec. Since these drives are not expected to be that reliable, the implication is that there is a reasonable chance of an error somewhere between 100TB and 1000TB for a drive rated <1 in 10^14.Of course in reality, other factors come into play. Personally, I chose RAID6 for my NAS. Spinning storage is cheap, so why not.
Robert Pankiw - Monday, May 22, 2017 - link
I think the error is a percent chance, which doesn't accumulate the same way as a "1 per xxTB written" would accumulate.It's probably something like (1-10^-15)^(10^15) which is about 37%, meaning that if you wrote about 113TBs (10^15 is about 113TB, unless you're one of THOSE people in which case it's 125TB) then you can expect about 1 URE. If you did that but didn't get a URE, and wrote another 113TBs, then expect the odds of not getting a URE to go from 37% to 13.5%.
ddriver - Monday, May 22, 2017 - link
I'd take ZFS over hardware raid any time.Sivar - Wednesday, May 24, 2017 - link
ZFS has issues of its own. For example, they require ECC system RAM for reliability to a far greater extent than most file systems. Performance can be much lower (because of ZFS's design, not because of hardware RAID performance). Operating system support is somewhat more limited. Of course, ZFS and hardware RAID are not mutually exclusive or even necessarily competitors.xplitz - Thursday, June 22, 2017 - link
That sounds stupid, the creator of ZFS pointed out that you don't have to use ECC for reliability, ECC is just "nice to have" not a requirement. You just wanted to say a bad thing about ZFS, I wonder what OS you use?spikespiegal - Wednesday, May 24, 2017 - link
URE's aren't the main problem with RAID5. Controller faults are the main problem with RAID5. Go to any server or enterprise focused forum and note the hostility towards RAID5, and it's well deserved.RAID5/6 essentially encrypts your data in the form of a parity stripe with very poor data correction unless you have a really high end controller. If the controller gets too warm, or has other issues the parity stripe can get corrupted causing a cascade of data integrity issues. Since RAID 6 creates two logical parity writes this acts as a form of fault tolerance which is why RAID6 doesn't get the ire of RAID5.
RAID 5 only had logical justification back in the 90's when a 9GB drive cost $700. Otherwise, RAID5 failures on servers is still my #1 form of data corruption. If you have't worked on enough storage to see the issue it's not my problem.
cm2187 - Tuesday, May 23, 2017 - link
If you care about your data you need a backup anyway. Then the chance of both your primary storage and your backup failing at the same time is pretty low even with 12 disks RAID.Sivar - Wednesday, May 24, 2017 - link
The impression that an unrecoverable error will necessarily be corrected by any RAID controller is not necessarily an accurate one.In the event of an individual disk error, RAID1 may be at greater risk because the controller has no way of knowing which copy is correct. RAID5/6 can derive a data mismatch through the parity information on other drives (not that all controllers necessarily do this).
One great thing about ZFS is that it stores checksums on parent blocks, allowing detection of data corruption regardless of the source.
dawp - Monday, May 22, 2017 - link
count me inedcoolio - Monday, May 22, 2017 - link
30 GB per $ ?! Awesome.I never thought I would see the day. Truly unbelievable yet inevitable.
cygnus1 - Monday, May 22, 2017 - link
The Helium makes all the platters easier to spin (ie use less power).
helvete - Tuesday, August 22, 2017 - link
What's the point? All of them are He-filled...SweeJ - Monday, May 22, 2017 - link
"The WD Pro 10 TB drive is engineered for personal or small business NAS systems with up to eight bays, is optimized for mixed workloads and has a 5400 RPM spindle speed. By contrast, the WD Red Pro 10 TB is aimed at medium business and enterprise-class NAS systems up to 16 bays...."Is this a typo? Shouldn't it be "The WD Red 10TB is engineered for personal...."?