Hey everyone! Hope you're having a great day. I'm reaching out to get some feedback on a project idea I have centered around K8S rightsizing. My background is in software engineering, and while I worked at a bank, I observed that a lot of teams are hesitant to use tools for rightsizing because they fear it could lead to underprovisioning, which may result in outages. So, I'm thinking of developing a tool that optimizes K8S clusters by favoring overprovisioning instead of underprovisioning. The idea is to make recommendations without actively scheduling changes. I have lots of features planned for the future, but I'd love to hear if this kind of tool would be useful for those managing K8S clusters. Would a solution that optimizes clusters while minimizing risk be of interest to you all?
2 Answers
It sounds like you're entering a crowded space. We already have some projects out there that aren’t quite working as expected. I'm building a cost-aware predictive autoscaler, and I’ve realized it’s complicated to classify resource allocation accurately. You'll need to consider various factors for predictions, like resource pricing and provisioning times. Plus, you might run into issues with tools like GKE's MPA. This asymmetrical optimization idea you're suggesting sounds a little unclear, though. Are you looking into something like the Descheduler? Just something to think about!
What’s your focus: node rightsizing or container rightsizing? Because the strategies might differ quite a bit depending on that. Just curious!
I focus on container rightsizing, but I can definitely extend it to node rightsizing!
Fair point — I probably explained it poorly. What I mean is: penalizing under-provisioning more heavily than over-provisioning in the recommendation algorithm. Similar to how Spot Ocean has an 'aggressiveness' slider. Does that make more sense?