클라우드 시스템 연구
Cloud Orchestrator
판단과 실행을 끝까지 추적합니다.
워크로드 데이터에서 출발해 권한 범위가 정해진 A–H 정책, 실행 검증, 독립적 결과 평가로 이어진 연구입니다.
처음부터 읽기이 프로젝트의 기록
이 순서로 읽어보세요
첫 문제부터 이어지는 설계와 개선 과정을 따라가세요.
Building a Causal Foundation for Failure Forecasting
From Borg trace windows to causal features, mature failure labels, and temporal validation: the data foundation behind Cloud Orchestrator.
Data Repair, Temporal Validation, and Two Different Prediction Tasks
Why causal schemas, mature labels, label-independent validation sampling, and model provenance matter—and why Borg failure forecasting differs from the live controller's next-window predictors.
Eight Roles, One Mutation: Designing the Control Boundary
The six-layer Cloud Orchestrator separates A–H responsibilities, feasibility, serial arbitration, durable execution, and independent assessment.
Tuning a Controller Without Tuning Its Evidence
Separating legacy PPO and Optuna experiments, local Kubernetes comparison, verified actuation, and a frozen independent CSC/CSP outcome gate.
Learning Without Surrendering Control
How Cloud Orchestrator combines eight specialist roles with an optional adaptive ranker, verified decision feedback, frozen comparisons, and evidence-gated policy promotion.