A few months after its beta launch, Google has announced the general availability of preemptible virtual machines as part of the Google Compute Engine cloud. Preemptible VMs have a lower price than other types of VMs that Google offers, but they can be shut down at any moment by Google with a 30 sec warning.
Beyond the possibility of being terminated, preemptible VMs have a few more limitations:
- their maximum continuous run time is 24 hours;
- they may not be always available;
- they cannot be live migrated or be automatically restarted after a maintainance event.
It must be also noted that preemptible instances do not affect the way prices are calculated for premium operating systems. E.g., if you install Red Hat Enterprise Linux and your instance runs for 35 minutes before being preempted, you will be charged a full hour. SUSE and Windows Server images, on the other hand, are charged by 1 minute increments, with a minimum of 10 minutes.
As Paul Nash, senior product manager at Google Cloud Platform, remarked,
What we’ve been doing between beta [release] and [this general availability announcement] is optimizing that balance and seeing how it works in production, so we don’t have to automatically just yank one away from someone, so we can keep that rate of interruption very low.
According to initial benchmarks, it appeared that in a pool of 50 instances about one instance was being terminated every five minutes. According to another early report, “the typical run time for a VM is about 10–20 minutes”. Those might well be early benchmarks, but Google has not provided any reference figure.
In a interview with InfoQ, Paul Nash clarified that when a virtual machine is preempted, it is simply terminated and its persistent disk is left intact, so data are not lost and work can be resumed later when preemptible capacity is available. Live migration to a non-preemtible VM is not supported yet, however.
In their announcement, Google also mention that during the beta period, a number of frameworks and products have been integrated with Google preemptible MVs, including Cycle Computing, a suite of products aimed to manage and scale large compute workloads which was previously only available on AWS, and Zync, a rendering solution.
Google preemptible VMs enter the same field as six years old AWS Spot Instances, which are characterized by a varying price based on a bidding mechanism. According to some relatively old estimation, Spot instances accounted for less than 5% of all AWS instances. No data has been released at the moment by Google about the relative penetration of their preemptible VMs, although, Nash wrote, “the system has seen extensive production use” with tens of thousands of jobs launched.