Azure Compute Fleet
Microsoft releases a new way to manage large compute estates
Last week Microsoft released the public preview of Azure Compute Fleet. No fan fare. No calls from my Microsoft HPC buddies. Not even a post on LinkedIn about it that I saw. If I didn’t diligently read release notes from the hyperscalers every week I would have completely missed it.
Microsoft, what is wrong with you!? Do you understand your customers at all? This is important for anyone running dynamic HPC workloads on Azure! It’s a potentially big deal. Did you use up your publicity quota on AI and pretending to fix your security problems 🤣
What were you going to do? Wait till it’s GA before telling anyone? Bad idea. You need to persuade them to adopt it early so they can tell you what you missed and what you got wrong. Because trust me… it needs some work. But I’ll come back to that another day once I’ve had more of a chance to play with it.
The TLDR is quite simply that you can just give Azure a list of SKUs (with a preference order), how many VMs you actually want in each spot and PAYG varieties and then just let it get on with it.
No more juggling multiple VMSSs. No more monitoring if you lost enough spot capacity to warrant spinning up some PAYG capacity to meet your SLAs. Well subject to a few limitations. It’s not quite as clever I would have hoped for.
