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 preemptible VM is “limited to a 24 hour runtime”. This is the maximum – it could be shut down at any time. You get all of 30 seconds notice.
Why would you want a VM like that? The idea is that Google gets to use spare capacity in its data centres, while customers get a generous 70 per cent discount, provided they design applications that can cope with the disappearance of some of its VMs from time to time.
A typical case would be where you need massive processing capability that is not time-critical. Another possibility is to use such VMs to enhance performance, while running the main part of an application on conventional VMs to maintain uptime.
Using preemptible VMs for load testing or security testing also makes sense, since the exact time when tests complete is unlikely to be critical.
Since data analysis is the most obvious use case for a disposable VM, Google has added an argument to its bdutil tool for creating Hadoop and Spark clusters, which lets you specify a percentage of preemptible VMs as part of the mix.
Smart cloud admins may also be able to create scripts that automatically replace preemptible VMs with conventional ones if too many of them are shut down, thereby reducing the cost of running a highly scalable application on Google’s cloud.
Amazon Web Services (AWS) already has a scheme for using up spare capacity at low prices. Spot Instances let you name your own price for a VM instance.
The higher you bid, the less likely it is that your instance will be “interrupted”, the AWS term for reclaiming its compute capacity by shutting down your VM.
Pre-emptible VMs are not a "new class of compute" after all then? Google says "unlike other clouds’ Spot Instances, the price of Preemptible VMs is fixed – making their costs predictable." Since you cannot influence the likelihood of termination by paying a bit more though, Google's offer is less flexible.
Along with preemptible VMs, Google has announced price reductions on all of its Compute Engine VMs, ranging from 30 per cent reduction for Micro instances to just 5 per cent for High CPU. The pricing model is complex, with a scale of automatic discounts when an instance runs for more than 25 per cent of the month. One thing is clear though: the cloud price wars are still going strong. ®