Data science has emerged as an integral part of enterprise business operations, aiming to uncover data-driven insights. However, volatility and variability are two major issues of these systems. The latter indicates the need to consider a multitude of conditions when assessing the accuracy of predictions made by analytics systems. This complexity stems from the understanding that each of these circumstances can alter at any given time. Consequently, the challenge is to estimate the predictive accuracy of data science systems even in the midst of uncertainty and variability. Such an estimate can enable good planning and potentially dynamic resource allocation for such projects. In this work, we present a study of various machine learning and deep learning models to estimate the performance of data science projects deployed in Apache Spark, a popular and flexible distributed analytics platform. Moreover, we create experiments for the purpose of gathering data, as well as assessing feature importance to find which input configuration contributes the most. We demonstrate the process of training such a model, from data collection to training and testing, and we systematically compare the various alternatives to help decision-makers choose the best one. Thus, by providing insights into the performance of data science projects under uncertain and variable conditions, this work offers valuable contributions to both research and practical applications to make informed decisions, plan effectively, and allocate resources dynamically in data science projects. Our results show that LSTM and MLP outperform other models in terms of response time models and throughput models.

Machine learning has quickly become an integral part of modern business. A lot of effort is put into developing and deploying machine learning models used in estimation, prediction of economic factors, or even in interacting with clients. This effort includes planning for tools and platforms to be used in training and validating these models, as well as allocating resources, time, and budget for these tasks. However, this planning remainslargely dependent on human acumen and is expensive to determine in a systematic fashion with automated tools. Benchmarking is the process of efficiently running experiments to determine a system's performance requirements, among others, in order to aid planning and resource allocation. Benchmarking intelligent and data-intensive systems remains in its infancy and does not cover fully realistic or very specific case studies. In this work, we propose SparkPerf, a benchmarking tool specifically designed for machine learning applications deployed with Apache Spark. SparkPerf focuses on longitudinal transactional workloads, which represent a more realistic class of case studies for enterprises, with high customizability, allowing users to test their own applications with their own, synthetically augmented, datasets. Our experiments demonstrate the benchmark's reliability, consistency, portability, and customizability.