Stream processing on demand for lambda architectures

Growing amounts of data and the demand to process them within time constraints have led to the development of big data systems. A generic principle to design such systems that allows for low latency results is called the lambda architecture. It defines that data is analyzed twice by combining batch and stream processing techniques in order to provide a real time view. This redundant processing of data makes this architecture very expensive. In cases where process results are not continuously required to be low latency or time constraints lie within several minutes, a clear decision whether both processing layers are inevitable is not possible yet. Therefore, we propose stream processing on demand within the lambda architecture in order to efficiently use resources and reduce hardware investments. We use performance models as an analytical decision-making solution to predict response times of batch processes and to decide when to additionally deploy stream processes. By the example of a smart energy use case we implement and evaluate the accuracy of our proposed solution.