Cloud Arbitrage Index (CAI): A Forward-Looking Risk Metric for Spot Instances

Abstract

Cloud spot instances offer considerable cost savings, but they introduce a major reliability challenge since they may be terminated when provider capacity tightens. Although cloud providers do show historical summaries such as price traces and overall eviction rates, they do not provide users with a forward-looking estimate of short-term interruption risk to guide workload placement. This thesis presents the Cloud Arbitrage Index (CAI), a predictive model for assessing short-term spot-instance interruption risk in Amazon Web Services (AWS). CAI is designed to estimate how likely it is that a one-hour spot-instance run will be interrupted if it starts within a particular six-hour window. To support this goal, the project starts by creating an active probing system that launches spot instances in different AWS regions and instance types in order to collect interruption data. It then builds a prediction pipeline that combines a statistical baseline based on recent pool-level behavior, where each pool is defined by a region and instance type, a machine-learning residual model, and a calibration step. The resulting risk estimates are evaluated as predictions and as inputs in downstream pool-selection policies that balance tradeoffs between cost and reliability. The results show that CAI performs better than both coarse statistical baselines and a direct machine-learning alternative that does not incorporate the baseline estimate on the main prediction task. Policy experiments further demonstrate that these predictions improve pool selection decisions under a linear price-risk objective. They also show that simpler prediction methods are competitive and can even outperform in retry-cost situations where local ranking among a small number of candidate pools is important. These results show that short-term spot-instance interruption risk can be predicted well enough to support cloud allocation decisions.