Note: the worn drives were very worn. 128GB drive with 280TB written is ~2400 cycles. >5x it's 480x rating!
Even though it's a cheap drive, it's rated endurance wasn't really that low. 600 cycles (1200TBW/2TB) is pretty common for consumer SSD: that's higher than 480x but not vastly higher.
Glad folks are chiming in with the temperature sensitivity notes. I have parts coming in for a much bigger home-cloud system, and was planning on putting it in the attic. But it's often >110°F in the summer up there! I don't know how much of a difference that would make, given that the system will be on 24/7; some folks seem to say having it powered on should be enough, others note that usually it's during read that cells are refreshed.
Doing an annual dd if=/nvme0n1 of=/dev/zero bs=$((1024*1024)) hadn't been the plan, but maybe it needs to be!
I would never buy a no-name SSD. Did it once long ago and got bit, wrote a program to sequentially write a pseudorandom sequence across the whole volume then read back and verify, and proved all 8 Pacer SSD's I had suffered corruption.
That’s also fairly common for cheap ‘thumb drives’, as I understand it. I’ve been bitten by that before.
(Edit: Allegedly if you use low-numbered storage blocks you’ll be okay, but the advertised capacity (both packaging and what it reports to OS) is a straight-up lie.)
I wouldn't say it's controversial but I suspect most people don't know about it. There's been a lot of discussion about SSD write endurance but almost none about retention.
This is a known issue. You have to power up your SSDs (and flash cards, which are based on even more flimsy/cost optimized version of the same tech) every now and then for them to keep data. SSDs are not suitable for long term cold storage or archiving. Corollary: don't lose that recovery passphrase you've printed out for your hardware crypto key, the flash memory in it is also not eternal.
The article states as much but to sum it all up as just that is leaving most of the good stuff out.
Perhaps the most interesting part of the experiment series has been just how much longer these cheap drives with tons of writes have been lasting compared to the testing requirements (especially with so much past write endurance on the one just now starting to exhibit trouble). Part of the impetus for the series seemed to be lots of claims on how quickly to expect massive problems without any actual experimental tests of consumer drives to actually back it up. Of course n=4 with 1 model of 1 brand drives but it's taken ~20x longer than some common assumptions to start seeing problems on a drive at 5x its endurance rating.
A not-so-fun fact is that this even applies to modern read-only media, most notably Nintendo game carts. Back in the day they used mask ROMs which ought to last more or less forever, but with the DS they started using cheaper NOR or NAND flash for larger games, and then for all games with the 3DS onwards. Those carts will bit-rot eventually if left unpowered for a long time.
I've noticed a number of GBA carts I've picked up used (and probably not played in a long while) fail to load on the first read. Sometimes no logo, sometimes corrupted logo. Turning it off and on a couple of times solved the issue, and once it boots OK it'll boot OK pretty much every time after. Probably until it sits on the shelf for a long while.
Please explain to me how is that supposed to work. For all I know the floating gate is, well, isolated and only writes (which SSDs don't like if they're repeated on the same spot) touch it through mechanisms not unlike MOSFET aging i.e. carrier injection.
Reading on the other hand depends on the charge in floating gate altering Vt of the transistor below, this action not being able to drain any charge from the floating gate.
According to a local expert (ahem), leakage can occur through mechanisms like Fowler-Nordheim tunneling or Poole-Frenkel emission, often facilitated by defects in the oxide layers.
If you at least read the data from the drive from time to time, the controller will "refresh" the charge by effectively re-writing data that can't be read without errors. Controllers will also tolerate and correct _some_ bit flips on the fly, topping up cells, or re-mapping bad pages. Think of it as ZFS scrub, basically, except you never see most of the errors.
It's a good idea to have a backup copy of the encryption keys. Losing signing keys is not a big deal but losing encryption keys can lead to severe data loss.
Endurance is proportional to programming temperature. In the video, when all four SSDs are installed at once, the composite device temperature ranges over 12º. This should be expected to influence the outcomes.
> - make one yourself by hacking the firmware: https://news.ycombinator.com/item?id=40405578
Be careful when you use something "exotic", and do not trust drives that are too recent to be fully tested
Do you realize the irony of cautioning about buying off the shelf hardware but recommending hacking firmware yourself?
That "firmware hack" is just enabling an option that manufacturers have always had (effectively 100% "SLC cache") but almost always never use for reasons likely to do with planned obsolescence.
Converting a QLC chip into an SLC is not planned obsolescence. It’s a legitimate tradeoff after analyzing the marketplace that existing MTBF write lifetimes are within acceptable consumer limits and consumers would rather have more storage.
Edit: and to preempt the “but make it an option”. That requires support software they may not want to build and support requests from users complaining that toggling SLC mode lost all the data or toggling QLC mode back on did similarly. It’s a valid business decision to not support that kind of product feature.
Tape is extremely cheap now. I booted up a couple laptops that have been sitting unpowered for over 7 years and the sata SSD in one of them has missing sectors. It had zero issues when shutdown.
While the tape is relatively cheap, the tape drives are not. The new ones typically starts at 4K USD, although sometimes for older models the prices can drop below 2K.
Tapes are cheap, tape drives are expensive. Using tape for backups only starts making economic sense when you have enough data to fill dozens or hundreds of tapes. For smaller data sets, hard drives are cheaper.
HDDs are a pragmatic choice for “backup” or offline storage. You’ll still need to power them up, just for testing, and also so the “grease” liquefies and they don’t stick.
Up through 2019 or so, I was relying on BD-XL discs, sized at 100GB each. The drives that created them could also write out M-DISC archival media, which was fearsomely expensive as a home user, but could make sense to a small business.
100GB, spread over one or more discs, was plenty of capacity to save the critical data, if I were judiciously excluding disposable stuff, such as ripped CD audio.
If you don’t have a massive amount of data to backup, used LTO5/6 drives are quite cheap, software and drivers is another issue however with a lot of enterprise kit.
The problem ofc is that with a tape you need to also have a backup tape drive on hand.
Overall if you get a good deal you can have a reliable backup setup for less than $1000 with 2 drives and a bunch of tape.
But this is only good if you have single digit of TBs or low double digit of TBs to backup since it’s slow and with a single tape drive you’ll have to swap tapes manually.
LTO5 is 1.5TB and LTO6 is 2.5TB (more with compression) it should be enough for most people.
> But this is only good if you have single digit of TBs or low double digit of TBs
That's not so enticing when I could get 3 16TB hard drives for half the price, with a full copy on each drive plus some par3 files in case of bad sectors.
Tape sucks unless you've got massive amounts of money to burn. Not only are tape drives expensive, they only read the last two tape generations. It's entirely possible to end up in a future where your tapes are unreadable.
After I was bamboozled with a SMR drive, always great to just make the callout to those who might be unaware. What a piece of garbage to let vendors upsell higher numbers.
(Yes, I know some applications can be agnostic to SMR, but it should never be used in a general purpose drive).
Untested hypothesis, but I would expect the wider spacing between tracks in CMR makes it more resilient against random bit flips. I'm not aware of any experiments to prove this and it may be worth doing. If the HD manufacture can convince us that SMR is just as reliable for archival storage it would help them sell those drives since right now lots of people are avoiding SMR due to poor performance and the infamy of the bait-and-switch that happened a few years back.
If you care about long term storage, make a NAS and run ZFS scrub (or equivalent) every 6 months. That will check for errors and fix them as they come up.
All error correction has a limit. If too many errors build up, it becomes unrecoverable errors. But as long as you reread and fix them within the error correction region, it's fine.
What hardware, though? I want to build a NAS / attached storage array but after accidentally purchasing an SMR drive[0] I’m a little hesitant to even confront the project.
A few tens of TBs. Local, not cloud.
[0] Maybe 7 years ago. I don’t know if anything has changed since, e.g. honest, up-front labeling.
Nothing can really save you from accidentally buying the wrong model other than research. For tens of TBs you can use either 4-8 >20TB HDDs or 6-12 8TB SSDs (e.g. Asustor). The difference really comes down to how much you're willing to pay.
Good question. There seems to be no way to tell whether or not we're gonna get junk when we buy hard drives. Manufacturers got caught putting SMR into NAS drives. Even if you deeply research things before buying, everything could change tomorrow.
Why is this so hard? Why can't we have a CMR drive that just works? That we can expect to last for 10 years? That properly reports I/O errors to the OS?
zfs in mirror mode offers redundancy at the block level but scrub requires plugging the device
> All error correction has a limit. If too many errors build up, it becomes unrecoverable errors
There are software solutions. You can specify the redundancy you want.
For long term storage, if using a single media that you can't plug and scrub, I recommend par2 (https://en.wikipedia.org/wiki/Parchive?useskin=vector) over NTFS: there are many NTFS file recovery tools, and it shouldn't be too hard to roll your own solution to use the redundancy when a given sector can't be read
Even though it's a cheap drive, it's rated endurance wasn't really that low. 600 cycles (1200TBW/2TB) is pretty common for consumer SSD: that's higher than 480x but not vastly higher.
Glad folks are chiming in with the temperature sensitivity notes. I have parts coming in for a much bigger home-cloud system, and was planning on putting it in the attic. But it's often >110°F in the summer up there! I don't know how much of a difference that would make, given that the system will be on 24/7; some folks seem to say having it powered on should be enough, others note that usually it's during read that cells are refreshed.
Doing an annual dd if=/nvme0n1 of=/dev/zero bs=$((1024*1024)) hadn't been the plan, but maybe it needs to be!
reply