Azure low-priority VMs become non-existent VMs, replaced by new Spot VMs

‘This workload will self-destruct in 30 seconds’ - now with variable pricing


Microsoft’s Azure cloud has started to offer Spot VMs, the pay-what-you’re-willing-to cloudy servers that hyperscalers use to sell off their unused capacity.

Microsoft has kind-of been here before with an offering called “low-priority VMs” that were sold cheap and terminated when Microsoft needed the capacity for a higher-paying punter.

Low-priority VMs had a set price whereas Spot VMs from the likes of AWS and Google vary depending on demand.

Azure’s new Spot VMs match that arrangement and also let you set the price you’re willing to pay for a VM. If Azure doesn’t have any VMs at the price you’re willing to pay, your workload won’t run. If you are running and the cheapest spot price exceeds the ceiling you’ve set, your workload will be “evicted” with as little as 30 seconds warning.

Which is why Microsoft recommends users “try to create checkpoints so that you can restart your workload from a previously known checkpoint to handle evictions and save time.”

Pricing appears to be keen: a Spot incarnation of a general purpose Av2 instance starts at $0.005/hour compared to $0.0075/hour for a PAYG instance and $0.0059/hour for a reserved instance on a one-year deal. Only a three-year reserved instance is cheaper than Spot instances, at $0.0038/hour.

Of course this sort of thing is not for everyone or every workload. Microsoft recommends it for batch jobs, disruption-tolerant workloads, stateless apps that need to scale and want to do so at lowest cost, short jobs or dev & test workloads.

Microsoft folk tell us Spot VMs are more flexible than low-priority VMs and are better at resuming workloads too.

Spot VMs are available in the entire Azure universe other than the Azure China 21Vianet region and on B-series instances. They can’t be used with promos and can’t “currently” use ephemeral OS disks.

The Register imagines Microsoft would not have launched Spot VMs unless Azure has the capacity to run them, which is good news by itself at a time when clouds have reportedly been rather busy coping with a flood of new coronavirus-created workloads. ®


Biting the hand that feeds IT © 1998–2020