Cloud systems research
Cloud Orchestrator
Control decisions you can trace.
From workload data to scoped A–H policies, execution checks and independent outcome evaluation.
Start readingIn this journey
The reading path
Follow the project from its first problem through the work that followed.
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.