김태오

The six-layer orchestrator was the point where the project stopped being only a data/model pipeline and became an actual control-plane experiment.I built it because I was tired of looking at model scores in isolation. A risk score is useful only if something can consume it. A demand estimate is useful only if it can affect efficiency behavior. Queue pressure is useful only if admission control c..

Once the data stopped feeling completely suspicious, I moved from the baseline forecaster to the advanced XGBoost track.The advanced track had a different purpose. The baseline forecaster was mostly a sanity instrument. The advanced XGBoost models were meant to become actual inputs to the orchestrator. That meant I cared about more than one prediction horizon, more explicit feature dictionaries,..

The first time the pipeline looked stable, I did not trust it. That instinct was useful, because the labels and schemas still had enough sharp edges to ruin the experiment.The main failure mode was not a loud crash. Loud crashes are easy. The dangerous failures were quiet: a nullable field interpreted too casually, a terminal event attached in the wrong temporal direction, a grouped trace row wi..

This project did not start from a neat ML idea. It started from being annoyed at Kubernetes behavior that I could observe, but could not really anticipate.At the company where I was working, I spent a lot of time around EKS, Kubernetes traces, GitHub changes, kubectl output, HPA behavior, and Karpenter behavior. Some days the cluster looked calm from far away, but the actual operational story wa..

After the initial forecaster idea, the project became mostly data plumbing. Not glamorous, but absolutely necessary.The first few runs taught me that trace work is less about training a model and more about refusing to lose track of what each row means. Borg data is not a single friendly CSV. I had usage windows, task events, machine information, optional columns, schema differences, and enough ..