JamesThurley on Mon, 12 Dec 2016 20:26:29
This is really just an open ended question which I was interested in hearing feedback on.
At the moment when an autoscale formula evaluation results in new nodes being requested, Azure Batch will create and image some new VMs and allocate them to the pool. This has the disadvantage of taking a few minutes between nodes being requested to nodes being available for use. Combined with the autoscale formula being evaluated only up to every five minutes, the time between submitting a large job and the tasks being executed at full capacity can often be between five and ten minutes.
Now don't get me wrong, Azure Batch is fantastic and has saved us a great deal of time and effort, and waiting that five to ten minutes is okay (and probably more than okay in many use cases).
But I'm imagining how responsive Azure Batch could be if it supported allocating nodes using containers rather than VMs. For our particular product and users, being able to go from job submission to scaled up in sub-30 seconds would have a big impact.
I guess the biggest sticking point would be that Microsoft would have to operate some kind of container-as-a-service infrastructure using a standard Windows or Linux image, which Azure Batch could tap into.
So my belated question is: Is this something that is being considered for Azure Batch, perhaps once Windows containers are generally available in Azure?
Fred Park [MSFT] on Mon, 12 Dec 2016 20:42:31
Thanks for the suggestion and feedback. If you want, please consider adding this as an idea here: https://feedback.azure.com/forums/269742-azure-batch
I know this won't be particularly helpful, but although we are investigating such scenarios, we have nothing to share at this point in time. Azure does offer managing clusters of container hosts via Azure Container Service and the Batch service does support 2016-Datacenter-with-Containers sku for use with Batch VirtualMachineConfiguration.
JamesThurley on Tue, 13 Dec 2016 11:16:11
Thanks Fred, I've added a suggestion here: