Service Level Agreements (SLAs) play a fundamental role in service compositions. As they represent binding contracts between service consumers and service providers, their violation is usually linked with profit loss. In order to maximize profit a multitude of SLA violation detection mechanisms exist that can roughly be classified into two categories: the ex-post and ex-ante detection, whereas the latter is fundamental when preventing SLA violations and subsequently is of great interest. Most of the current approaches either focus on the ex-post detection or insufficiently consider the behavior of SLAs over time.
Moreover, the accurate prediction of target values, being of categorical nature, is still open. The following thesis elaborates a new ex-ante SLA violation detection mechanism, which makes use of the general statistical class of Autoregressive Integrated Moving Average (ARIMA) models. Thereby the past realization values of specific target values - called Service Level Objectives (SLOs) - are analyzed and fitted into a time series model. As the application of SLAs involves the validation of as many services as possible at the same time, the reduction of human intervention is desirable. Therefore, the creation and training of the statistical models is done automatically. All theoretical models are implemented and tested within the research prototype Vienna Runtime for Service Compositions (VRESCo) and the results are validated in the course of a practically relevant example use case, proving the great potential of ARIMA models in the domain of forecasting SLA violations:
Differently behaving SLAs (e.g. the delivery time of a reseller) result in a prediction accuracy ranging from 99.31% up to 99.77%. In the case of categorical SLAs the resulting prediction accuracy of 7.18% indicates the need of further research work.