It started innocently. We should add auto scaling. This is crazy. We're doing manual scaling. Well....

Six months later, we had deployed HPA, Cluster Autoscaler, Karpenter, KEDA, VPA, AND Kueue. Each one solved a problem and created three more. This is that story: the dumb failures, the "why didn't we read the docs" moments, and the gotchas that only show up in production.
5 minutes. 6 autoscalers. A whole lot of regret. An object lesson in what not to do when dealing with auto scalers, even as we evolved into AI workloads
I'll speed-run through each autoscaler in the order we actually adopted them... HPA + Cluster Autoscaler together on day one, Karpenter when CA was too slow, KEDA when CPU metrics failed us, VPA when we finally admitted our resource requests were fiction (because why not), and Kueue when AI training jobs started fighting each other.
Come laugh at our pain. Leave knowing which autoscaler you actually need or at least which one you want to avoid, and which gotchas will bite you.