Why Google Wants Hard Drives To Be Less Reliable

The rise of cloud storage means that individuals and corporations are buying fewer hard drives, while the operators of cloud services are purchasing them in massive quantities. Now Google is arguing that as a consequence, hard drives really need to become less reliable — and that's not quite as insane as it first sounds.

Disk drive picture from Shutterstock

Google outlined its vision of the future of hard drives in a recent blog post by Google's infrastructure VP Eric Brewer and an associated white paper released at the same time.

Unsurprisingly, Google is a massive consumer of hard drives in its data centres. The volumes of storage Google needs are insane: as the post notes, YouTube alone requires a petabyte of new storage every single day, thanks to the 400 hours of video which are uploaded to the service each day.

Most Google customers don't end up paying for that storage either, though Google is obviously bringing in enough cash for that not to be a concern. (Last year, it pulled in $US74 billion in revenues.)

So while individuals are increasingly relying on SSDs and USB storage, or simply uploading everything into the cloud, Google (and its rivals such as Amazon and Microsoft) are soon going to be the biggest buyers of drives, as Brewer notes:

The rise of cloud-based storage means that most (spinning) hard disks will be deployed primarily as part of large storage services housed in data centers. Such services are already the fastest growing market for disks and will be the majority market in the near future.

That's good news for hard drive manufacturers, since they'll still have somebody to flog their wares to. But Google makes a good case for needing to rethink the current approach to drive manufacture and design.

The main reason is because current hard drive designs don't reflect the data centre use case. One simple and obvious example: hard drives maintain their current form factor because they needed to be roughly the same size as the 3.5-inch floppy drives which preceded them. This is undoubtedly ancient history to most Lifehacker readers, but in the late 1980s and early 1990s, a hard drive wasn't a given in a PC. Keeping the same form factor as the then-predominant form of storage meant that those who could afford the luxury of a hard drive had the option, but the system design didn't completely need to change. That made sense in 1989, but it isn't necessarily sensible a quarter of a century later.

More crucially, data centre operators aren't particularly concerned about the reliability of a single disk. The fact is that all hard drives will fail eventually, and any data centre system will ensure that data is stored on multiple drives so it can be restored when a failure happens. As a consequence, we don't necessarily need to focus on improving the reliability of drives beyond current levels, Brewer argues:

We need to optimise the collection of disks, rather than a single disk in a server. This shift has a range of interesting consequences including the counter-intuitive goal of having disks that are actually a little more likely to lose data, as we already have to have that data somewhere else anyway. It's not that we want the disk to lose data, but rather that we can better focus the cost and effort spent trying to avoid data loss for other gains such as capacity or system performance.

So what might those changes look like? They could include altering the physical format of disks to make them taller, or changing the firmware in hard drives to offer a higher rate of I/O (even if it's slightly less reliable and predictable than previous versions).

Height might not seem like a major factor in drive storage, but as the paper explains, it could substantially increase performance:

We propose increasing the allowable height ("Z height"). Current disks have a relatively small fixed height: typically 1" for 3.5" disks and 15mm maximum for 2.5" drives. Taller drives allow for more platters per disk, which adds capacity, and amortises the costs of packaging, the printed­circuit board, and the drive motor/actuator. Given a fixed total capacity per disk, smaller platters can yield smaller seek distances and higher RPM (due to platter stability), and thus higher IOPS, but worse GB/$. The net result is a higher GB/$ for any specific IOPS/GB that could be achieved by altering any other single aspect, such as platter sizes or RPM alone.

Google isn't proposing an exact height or any other design parameters. As the paper notes, while Google buys enough disks that it could create any specification it liked and create enough business for a manufacturer to sign up, it would prefer a universal standard. That might require some delicate negotiations with the other major cloud providers, but it's a sensible goal.

Those changes won't ever be seen in consumer drives, but they would make a big difference to cloud service operators. And while they won't come overnight, Google has sufficient clout that its vision of the future of the hard drive is likely to bear fruit.

Angus Kidman is editor-in-chief for comparison site finder.com.au </a, a former editor for Lifehacker Australia and a man who still has the first 5.25in floppy disk he ever purchased. Follow him on Twitter @gusworldau.


Comments

    YouTube alone requires a petabyte of new storage every single day, thanks to the 400 hours of video which are uploaded to the service each day.Not quite. YouTube users upload 400 hours of shi... um... video every minute.

      Its so much S*** they upload every minute, that I wish there was a way to have them deleted...

    Fair point. The easiest way of adopting the recommendation would be for Seagate or WD to start making drives that are a multiple of the standard height of a 3.5" drive. (so they're twice or 3x as tall). Then they'd fit into existing cases / enclosures with the minimum of re-engineering.

      The original 3.5 inch drive was twice the height of the current form factor anyway.

      My first 3.5 inch external drive for my old Amiga was not much smaller than an actual brick. However, that did include the external case.

    I'm not sure how I missed the return of Angus Kidman...
    Careless checking of bylines :(

    Cloud storage is viable option only if you have decent Internet connection speeds. With 1 Mbit uplink connection it is just too slow. So far it's good as data backup.

    Bring back 5 1/4 inch drives?
    Most PC's still have bays of that size, they just use them to hold CD/DVD drives.

    Since you are resurrecting an old format, you can even have full height drives that take up 2 bays in height.

      Yep, I remember 5 1/4" full height drives from the mid-80's. 10MB.

        I think you could fit more into a drive of the same physical size these days...

    It's almost a contradiction in terms. If the major market for hard drives is fast becoming the cloud rather than consumers, then it makes sense to increase not decrease the reliability of drives. Why? Simply to keep the massive cost of replacement down. It doesn't make commercial sense to reduce the reliability and increase the overall costs of replacements unless the form factor is changed significantly to bring the overall cost per unit of storage down a lot to compensate for the lower reliability.

    At the same time, more and more people use either external USB drives, or NAS devices to store large volumes of data, and small sized SSDs in laptops. So, it also makes sense for drive manufacturers to change the form factor for consumers as well, because nobody cares as much these days about squeezing a hard drive into a PC case.

      I can tell you that a data centre using RAID5 for a disk array gets very, very nervous after the failure of the first drive. We wound up standardising on RAID6 (or in some cases, RAID10) in the end.

      Talk about drive reliability being unimportant ignores the SMB market, where drive reliability is absolutely critical. If you have LOTS if storage redundancy it doesn't matter, but that level of redundancy is very expensive in terms of supporting hardware.

        When a drive fails, until it is replaced you are vulnerable to data loss. That's why a lot of people use RAID-10 / SAME (stripe and mirror everything) as their default choice. Individual drives could be then be less reliable, in theory, but the cost of a worker replacing the drive can be more expensive than the drive itself. I'm sure Google would rather buy more reliable drives for a few dollars more rather than pay the costs of someone replace their drives — given the number that they are likely to have it would almost be a continuous process. I'm wondering, Angus, have you ever managed or work with IT at scale?

          Not sure why you called me Angus, especially when we seem to be pretty much on the same page where drive reliability is concerned.

          Anyway, the staffing costs of replacing drives can be mitigated to some extent by having hot spare drives in place. However, there is still a window of vulnerability where a drive has failed but redundancy has not been restored. It's that window that gives IT managers nightmares.

          Generally I'd go for RAID10 where R/W speed was needed, RAID5 where minimising drive waste was important but losing the data would not be utterly catastrophic (mainly, when backups are relaible) and RAID6 (with hot spares, if possible) where redundancy was utterly vital.

          The problem with RAID10 is that when one drive fails, there is still a small chance that a second drive failure will affect the single drive that has lost redundancy before the RAID rebuild is complete. Or that the hot swap spare may be fauilty.

          This has in fact happened to me once. That was not a happy week. (Fortunately we had fairly comprehensive backups, but with several terabytes of live data - and some "noncritical" systems not being backed up that had inadvertently become critical - staff and management were Not Happy.)

            I think you're missing the point - Google doesn't use RAID to protect disks. Its uses the filesystem to redundantly distribute files across all of the storage devices.
            https://en.wikipedia.org/wiki/Google_File_System

            Yes you can build in drive hot spares - just remember its not about rebuilding a drive in a RAID array. Its about the replacement disk picking up on some of the redundancy.

            You can amortise labour costs by doing bulk replacements... and I guess swapping out a tray (or rack) of storage as > x% of drives have failed is pretty cost effective.

              Of course I understand that's how Google does things. My point was that while reliability was increasingly unimportant for Google, there are still a lot of people for whom drive reliability matters.

              Anyway, for drive replacement, as long as you have hot swap drives, having somebody wandering through the data centre with a cart full of spare drives, tossing the old ones, is not all that expensive. Swapping a drive takes maybe a minute. Even if the drives are well separated, you're probably looking at just a few minutes moving between drives, so only a couple of dollars of staff time per drive in doing your drive replacements.

              If Google halves the reliability of their drives and in so doing cuts cost by a third, it would be well worth them doing so, assuming they are using other means of improving redundancy.

              After all, RAID6 only costs two drives out of your drive pool. If you have a thousand drives, and sufficient smarts to cut overheads from parity rebuild when there's a drive failure, there's no reason you couldn't have some flavour of RAID6++ where 5% of drives are dedicated to redundancy and the statistical chance of failure is less than the chance of Tony Abbott winning the next election.

              Not everybody runs thousands of drives, however, leaving the options of hiring space from one of the big players or else having a local installation with traditional "reliable" hard drives. So unreliable drives may be good for Google, but that doesn't make unreliable drive in general to be a desirable thing.

            Angus being the author of the article.

              I saw that, but wasn't sure why you were responding to my post while apparently addressing your message to Angus.

Join the discussion!

Trending Stories Right Now