An empirical study on predicting cloud incidents |
| |
Institution: | 1. Computer Information Systems, Eastern Michigan University, 419 Owen Bldg, Ypsilanti, MI 48197, United States\n;3. Decision and Information Sciences, Oakland University, 342 Elliott Hall, Rochester, MI 48309, United States |
| |
Abstract: | With the increasing rate of adoption and growth of cloud computing services, businesses have been shifting their information technology (IT) infrastructure to the cloud. Although cloud vendors promise high availability and reliability of their cloud services, cloud-related incidents involving outages and service disruptions remain a challenge. Understanding cloud incidents and the ability to predict them would be helpful in deciding how to manage and circumvent future incidents. In this study, we propose a hybrid model that employs machine learning and time series methods to forecast cloud incidents. We evaluate the proposed model using a sample of 2261 incidents collected from two cloud providers namely, Netflix and Hulu. Unique to this study is that our model relies solely on historical data that is independent of the underlying cloud infrastructure. Results suggest that the proposed hybrid model outperforms individual forecasting models: neural network, time series and random forest. Results also reveal important temporal insights from the proposed model and highlights the practical relevance of historical data to forecast and manage cloud incidents. |
| |
Keywords: | Cloud services Cloud incidents Hybrid model Forecasting |
本文献已被 ScienceDirect 等数据库收录! |
|