Hacker News Reader

SSDs are meant for boot and application drives where lots of random I/O happens. You wouldn't normally use them in RAID except to RAID0 them for better performance and more space.

Instead of anecdotes about failure rates, it's best to look at data. I don't know of any large-scale studies of SSD reliability, but http://forums.storagereview.com/index.php/topic/29329-ssd-fa... looks at return rates, which correlates with failure rates. Intel's return rate is 0.6%. The other SSDs are 2-3%. Hard drives vary from 1-6% for 1TB and 3-10% for 2TB drives. SSDs are more expensive than hard drives, so their return rates should be higher for the same failure rate (people are more likely to return a broken $300 SSD than a $50 hard drive).

The reason many people like SSDs is because they widen the worst bottleneck in modern computers (random disk I/O) by a factor of 10-100. The difference is very noticeable. I highly recommend you try out an SSD as your boot/application drive.

> SSDs are meant for boot and application drives where lots of random I/O happens. You wouldn't normally use them in RAID except to RAID0 them for better performance and more space.

Most folks who really care about massive I/O performance are OLTP implementers (i.e. big Oracle/Postgres/MySQL installations). Those people also really care about reliability, so as long as the failure rate of SSDs is non-zero, even with screaming random IOPS capability, they'd be fools not to use some RAID 1 or RAID 5 implementation.

I'm not entirely certain return rates are a good indicator of failure rates.

First, you only return a hard drive if it's under warranty so that gives you a 1-2 years window.

Second, if you care about security, you don't really return hard drives. Once bought, unless you are certain everything is encrypted on them, the drives never leave the premises.

You can usually get a deal with hw/drive vendors to train and certify your in house staff to declare drives dead and then field destroy with replacement credit.

Intel's published reliability data shows a ~0.4% annual failure rate of the X-25M: http://images.anandtech.com/reviews/storage/Intel/320/reliab...

They claim greater reliability for the new 320 series which is based on the same controller hardware, and have upped the warranty of that from 3 to 5 years: http://newsroom.intel.com/community/intel_newsroom/blog/2011...

There definitely has been a lot of noise about high failure rates in OCZ's Vertex series, but it doesn't seem to be anything inherent in SSDs or even the SandForce chipset. OCZ shipped SSDs with pre-release firmware, uses a number of different flash chip vendors, and may not test sufficiently prior to release.

I'd say buying a reliable SSD is more difficult than buying a HDD since there are quite a few variables, but there is sufficient information available these days to make an informed choice.

Performance is really worth it though! I wish there wasn't such a focus on boot times because I only have to reboot my MacBook Pro for certain OS updates or when it crashes, but high performance random IO makes a huge difference in day to day use. Applications launch almost instantly, I can do a find+grep on nested directories containing thousands of files in a fraction of a second, and compile times are pretty much only limited by CPU speed.

Anybody have some detailed technical insight into why the failure rate has been so high? (i.e., I can guess as well as the next guy, I'm asking for knowledge.) One of their initial promises was reliability due to not having moving parts, and, well, that sounds pretty reasonable as far as marketing copy goes. Why isn't it true?

jerf, the oddest part is that if you read through the de-duplication and redundancy designs of modern SSD controllers (like the SandForce SF-2k series) not only do the drives do automatic de-duplication and compression on the fly, the scatter the bits evenly across the memory arrays (in parallel) to control even wear-and-tear of R/W cycles on the memory.

Then they typically write the bits to multiple locations or an entire section of the drive is maintained for parity information, sort of like an on-disk RAID5 setup where the individual blocks of NAND can be though of as "disks". For example, a 120GB drive typically has something like 150-160GB of storage space on the disk. This is not unlike what spindle disks do with a large section of disk not allocated to the format operation and kept behind for failed sectors that are re-allocated by the controller.

So all that work and the failure rates are still too high for comfort. It's really surprising to me and suggests that all of this needs more baking time.

I have no idea what the failure rates are in lower-end drives like the Kingstons and other drives that are likely rebranded garbage. I imagine it's not any better :)

the interesting this is that Macbook Air are all SSDs - so unless OSX is doing something terribly creative, why arent we seeing the same kind of return/failure rates on Macbook Airs ?

There's a tradeoff between speed and reliability, and the Samsung SSDs that I believe are used in the Macbook Airs lean towards reliabilty at the expense of having non-random IO no better than a normal hard drive. People who buy SSDs by themselves, on the other hand, are often just going down a list of offerings comparing IO speed to price and not having any good way to look at the reliability.

I haven't seen that they aren't.

That, and you never own a MacBook Air more than a couple years, because Steve Says To Buy A New One Then.

chuckle I was expecting the indignant Apple downvotes =)

They fail (or rather you hear about them failing) because they are new. I have graveyards filled with spinny disks that failed, but it's hardly newsworthy because everybody knows drives fail. Has anybody put half HDs and half SSDs in a data center of any size? Anecdotes and data and all that.

It is a physical quality of flash memory that each flash cell has only a limited number of writes it can handle before it fails. If you want to know the dirty details, what happens is that electrons eventually get trapped in the dielectric of a flash cell and ruin everything. This cannot be fixed unless someone invents a new electron capture proof dielectric.

Flash memories use all kinds of electronics to ameliorate this problem, but in the end most they can do is try to equalize the number of writes each cell experiences, which results in the drive working at close to its full capacity for as long as possible but then failing very quickly afterwards.

Mind you, there have been people working on clearing up trapped charges. I doubt there's been much usable progress (IIRC, the methods involve some kind of irradiation of the dielectric, which is difficult to perform in situ.)

If this was the failure mode experienced by most drives that fail, though, it wouldn't happen for years and even after you were no longer able to write to the drive you would still be able to read from it. This might be how the more reliable Intel or Samsung drives end up failing, but all evidence is that the failure modes of the highest performance but less reliable drives seem to be more interesting than that.

Why do you think it would not happen for years? As I mentioned in a previous post there is a trade off between price memory size and failure rate so some of the higher density memories will have much lower life than what you may be used to.

And no you would not be able to read from it. When a flash cell wears off the ability to read it degrades or disappears. Depending on the memory you still may be able to recover the data using ECC or the like, but this ability tends to disappear as large numbers of cells start dying.

I have had the pleasure and challenge of learning a lot about flash memory and I must add that the high failure rates are not aberrations, they are more or less expected. When people talk about SSDs they assume they are like hard drives. I.e., they should last several years, unless there is some kind of defect on them.

Well, that is not true. Any flash memory cell is basically limited as to the amount of times it can be written to. This is the result of certain physical properties of the materials used, and cannot be avoided.

This means that any SSD has a limited number of writes on it before it fails. So SSDs will fail and ones that are written to a lot will fail faster. If you use an SSD for virtual memory, for example, you can really speed up your computer (if it does use a lot of virtual memory) but your SSD will fail quickly.

There is generally a three way trade off between cost, storage size and failure rate. So every company would probably strike a slightly different trade off, so it is not surprising one company's drives to last longer than another's. Furthermore, within a single type of drive different times of failure would mostly depend on how much the drive is written to.

In general we should not accuse companies who make drives that fail of making shoddy merchandise (unless they fail way too soon), this is just the nature of flash memory.

No. The firmware on SSDs utilises wear leveling (they also set aside a certain percentage of flash cells for housekeeping), so for the vast majority of people for whom SSDs have failed, the reason is NOT that the flash cells have died. From what I read, the real reason seems to be software bugs in the firmware.

I can't find any refernces on this. Do you have a link you can share?

Fargren,

Actually all SSDs have to do this. The failure rate on NAND memory is so high at the higher densities that are being used now, that it would be impossible to ship a drive that didn't do it.

You can learn more about it here[1] - while that is a review of the Vertex 3 (SandForce SF-2500 based SSD) the approach (de-duplication, write load-balancing, etc.) is same for all the controllers out there.

There is another awesome article Anand did explaining all the features of some of these higher end controllers and how much work they actually do, but for the life of me I cannot find it. I'll follow up if I do.

[1] http://www.anandtech.com/show/4159/ocz-vertex-3-pro-preview-...

Was it this:

http://www.anandtech.com/print/2738

?

Damn close, but not quite. That article Anand put out during the height of TRIM-gate after discovering the permanent speed-degradation designed into the early drives as they filled up (fixed in all newer controller designs FWIW).

The one I had in my brain was one done after that, it was focused on the SandForce SF-2k series (might have been the 2500) where Anand goes through all of the features in the controller pertaining to data safety. The way it is explained and marketed makes it sound like an SSD will fail once every 15 years, but in reality their failure is horrible... I have no idea why the massive discrepancy. Either these NAND chips are MUCH more volatile than we are being told and Anand's numbers are based on (he has calculations to determine failure rates based on load balancing) or the data parity/recovery systems in these controllers are so insufficient that they underestimate the actual data loss occurring on these things and thus underperform.

Thanks for trying to dig it up though, that article along is certainly worth a read for anyone interested in this topic.

I am very well aware that a flash memory performs wear leveling. However this does not prevent the flash cells from failing, it merely ensures they all fail as late as possible. Wear leveling does not prevent wear it only ensures that wear is generally even between cells.

I'll take reliability over almost anything. I work out on the road 200+ days of 365. A drive failing on me while out in the wilds would be a ~~~~ing catastrophe. I can picture it now, Sunday night, after a long flight, code modified on the flight, not in a remote GIT update in a couple of days... failure in such a scenario would mean i'm out of job. No thanks, I'll take the slow HDD

At my last job, they got me an SSD. About a month later, I started coding a new feature for their product. It was a few hours' worth of work, and they were using CVS, and didn't really have a good branching model, so I didn't do any work-in-progress checkins.

While I was coding the feature, the SSD failed. This manifested itself via all sorts of strange compiler errors, and so I rebooted. After that, Windows didn't show the SSD anymore, and my code was never seen again.

So, while the SSD did save me some time, I effectively lost it all when it failed, due in part to my stupidity, but also due to the fact that you need to understand that if you buy an SSD, you MUST plan for failure. This is not an option, this is "yes-I-will-wear-a-helmet-while-riding-a-motorcycle-at-70-mph".

Contrast that with my Dell laptop. I bought it in ~2007, and I've never had a problem with it (except I've had to replace the dim screen a couple times).

If you buy storage media, you must plan for failure. Having worked in data centres, I've seen spinning disk drives DOA, die after a week, die in clusters (sequential serials, different machines), etc.

SSDs aren't 100% reliable, but neither are HDs. You lost your SSD at an inopportune time, and that sucks, but it could just as easily have been a spinning disk that died, or it could have died after you finished a nightly system backup so you had no data loss at all. It's pure coincidence.

> I'll take reliability over almost anything.

Then measure reliability by IO operation instead of wall clock time.

I had a catastrophic failure in my MBP HDD after a couple years of use so heavy that if it were an HDD, I simply wouldn't've been able to have done all that work.

HDDs do fail, too. A lot. If drives failing for you is a catastrophe, you should consider taking backup media with you.

This may be anecdotal, but I've never had a HDD fail and neither do I know anyone who had a HDD fail.

And how many drives do you have, and how long have you had them? In the last 15 years I've probably had 20 HDD's fail out of 150 total. There was a particularly bad period where I was dealing with about 15 Dell laptops with 1.8" HDD's. Many of those drives failed, some repeatedly. Bad enough that I was buying 5 drives at a time until we could replace the computers.

HDD's from every manufacturer can and do fail at fairly high rates.

If I'm counting correctly I've had 11 drives, and each was used on average 3-5 years I think. I've helped a lot of other people with their computers being the nerd kid, and the problem has not been a failed drive yet (and it has been other failed hardware, like RAM and video cards and sound cards).

If you've had about 20 drives fail, and of 15 Dell laptops a lot of drives failed, some repeatedly, don't you think that your bad experience was primarily caused by that type of drive?

Hard drive failure is probabalistic. You're just one data point. There are plenty of people who have seen a moderate number of drive failures, and some people who have so many drives fail that you have to check that they aren't tapping them with a magnet for good luck or something.

Not believing in hard drive failure because you haven't had any, despite the rest of the word telling you it's a fact of life... well, it says something interesting about human psychology, is all.

Sure I believe in hard drive failures. I'm not saying that hard drives don't fail, I'm just providing my experience with hard drive failure. Where did you get the idea that I don't believe in hard drive failure?

I realize that I'm just one data point. What I don't get is why people downvote me for providing this data point, whereas people don't seem to have a problem with blanket statements like "HDDs do fail, too. A lot." that provide zero data points.

Moreover we are talking about SSDs vs HDDs here. "HDDs do fail, too. A lot." may be true, but if according to Jeff Atwood, a SSD proponent, on the order of half of SSDs fail in the first year then HDDs may fail a lot, but SSDs fail a whole lot more.

I've had a 256GB Crucial SSD running for a year without issue. I wrote about it at http://littlebitofcode.com/2011/03/02/a-year-with-a-ssd. I'm not concerned by failure rates because I have an excellent backup strategy. I've had dozens of spinning disks fail unexpectedly, they can and will fail as spectacularely and unexpectedly as an SSD could.

After 1 year Crystal Reports is giving you a 67% health rating? That seems... not great. I don't know how full the drive is, but I imagine you'll be look at a failed drive by 1.5year mark. For $630 I'm having a hard time getting excited about it.

Totally agree that SSDs will fail as big and randomly as spindle disks, but for 10x the price (or whatever factor it is) is it wrong to expect better reliability?

If you need the speed and have the money, more power to you, especially if you have a good backup strategy and losing a disk is no biggie... I just don't think many people realize HOW unstable these things are. That is all I'm pointing at.

The warranty is five years. I won't lose any data, and in two years I'll get a new drive from the manufacturer. I don't expect any reliability, hard drives aren't reliable, this is nothing new.

The Crystal Disk Reports is based on the flash wear patterns, so if it's at 67% it is likely to last 2 more years not six more months. The way I calculated, the drive saved me enough time in the first year to pay for itself.

It cannot be stressed enough that hard drives (of any type) are not reliable. If your storage strategy is predicated on the assumed failure rates of your hard drives then your in for a nasty shock down the road. And as stated proper warranties eliminate most of the short term risk associated with failures.

I've lost 2 Crucial drives in 2 years.