back to article Psst. Want a cheap cloud, VM? Google has one. But there's a catch

Google has announced what it calls "a new class of compute": virtual machines (VM) that are predictable in price but not in operation. "Preemptible VMs are the same as regular instances except for one key difference – they may be shut down at any time," says senior product manager Paul Nash. Another restriction is that a …

  1. SecretSonOfHG

    Unpredictable response times at best, never ending jobs at worst?

    If Google starts shutting down your VMs and there's no "bid" process in place like AWS, one has to assume that it does so because there's another customer paying more that wants the capacity or it is Google itself that needs the resources. That means that you won't know when they will become available again, if ever.

    So what Google is selling to their customers is the idea of setting up some higher price, always available VMs as a kind of minimum working set and then having the option to pay more for their jobs to run faster, only they can't tell them how much or how fast they will go, the only thing being for sure is that the faster they go, the more you pay.

    Not sure how many customers are willing to go for that schema. I can imagine only a few uses (one-off data migrations or conversions, test runs and what else?) where this makes sense, but perhaps I'm not experienced enough.

    1. the spectacularly refined chap

      Re: Unpredictable response times at best, never ending jobs at worst?

      Not sure how many customers are willing to go for that schema. I can imagine only a few uses (one-off data migrations or conversions, test runs and what else?) where this makes sense, but perhaps I'm not experienced enough.

      I can see plenty, so many in fact that there's a special term for it. Not some newfangled buzzword from a marketroid but a real term that actually means something - good old fashioned batch processing.

      I'll admit that I'm struggling to see applications here that would fit well but that is more a reflection of our infrastructure than the merits of the offering - anything here is either not substantial enough to justify configuring a VM for or the amount of data that needs uploading is out of proportion to the CPU time requirement. However, we do most of our processing in house, very little cloud which is essentially directly customer-facing stuff. If that wasn't the case and our data was already cloud based then something like this would be very attractive for many tasks.

  2. Charles 9 Silver badge

    If you read the article itself, you'll note they put up some use cases: mostly computer-intensive but non-critical operations. If they don't finish, oh well, pick up again later. As for who's usurping your VM, I think in this case it's Google itself.

  3. sysconfig

    Plenty of use cases...

    - batch processing of all sorts

    - auto-scaling (just launch two at a time when you only need one, so that it doesn't hurt if one goes down, and still pay less), obviously much more efficient and feasible at larger scale

    - automated testing

    - etc etc

    The problem with Google's offer in comparison to AWS's Spot Instances is that it's too expensive and limited to 24 hours max.

    In AWS you regularly get spot instances for below 20% of the normal price (just tested with a m3.large instance for $0.03 per hour, instead of the normal $0.154 per hour; became available within the normal time frame... two minutes).

    That said, prices usually spike dramatically when there's high demand, which means it almost doesn't matter how much you bid. If the spike kicks in, your instance goes down.(Spike prices are often higher than the regular instance price, so unless you bid insanely high amounts, your instance is going to go away.) That's fine though, because you get serious compute power for very low prices, and you know it's not going to last. Design your application to cope with it and you can actually save quite a bit.

  4. NoneSuch Silver badge

    Pirate Bay might want a bit of irregular processing.

  5. jr424242

    Replace pre-emptible instances?

    I get the part about running batch stuff off line, thanks for the discount, it's great.

    However, why would I expect to be able to spin up a full-price instance immediately after my pre-emptible instances was shut down? The pre-emptive instance was shut down for lack of capacity at some level; if everyone tries to spin up an equivalent paid instance, there is still no more capacity. (With a 70% discount, some bright spark will automate all this for MapReduce work, so in fact everyone *will* be doing this, but it won't scale.)

    Auction pricing for spare capacity would work better. I wonder if Google could implement an auction at that scale, it might have other business applications :)

  6. Nate Amsden

    not VMs

    These are containers not VMs right?

POST COMMENT House rules

Not a member of The Register? Create a new account here.

  • Enter your comment

  • Add an icon

Anonymous cowards cannot choose their icon

Biting the hand that feeds IT © 1998–2019