A weighted sensitivity metric for predicting latency in a KAFKA cluster

The subject of this article is sensitivity analysis methods used to assess how variations in input parameters affect the output results of a model or system. The aim of the study is to develop a new approach to sensitivity analysis that combines classical parameter impact assessment methods (Morris and Sobol methods) with metrics that capture structural changes in the distribution of output data (Euclidean distance, Hellinger distance, Jensen divergence). This approach allows for evaluating the influence of a parameter not only in terms of the amplitude of its effect, but also in terms of changes in the shape and structure of the probability distribution of the results. To achieve this objective, the article addresses the following tasks: formal definition of a new sensitivity analysis approach; development of a Bayesian network for modeling end-to-end latency in a Kafka cluster; performing sensitivity analysis using the proposed approach; and conducting an experimental study using the calculated parameter influence weights to initialize the weight matrix of a neural network that predicts Kafka cluster latency based on selected configuration parameters. To accomplish these tasks, the study applied methods from the theory of experiments, Euclidean geometry, statistical distribution theory, information theory, machine learning, Bayesian statistics, and graph theory. Results: To evaluate the effectiveness of the proposed approach, comparative training of a neural network was conducted using various weight initialization strategies. Analysis of the loss function, constructed using the mean squared error minimization criterion, showed that the lowest values were achieved by the model initialized with weights obtained using the proposed parameter influence estimation approach. Conclusions: The study proposes a novel approach to sensitivity analysis. The innovation lies in integrating the strengths of both causal-oriented and variance-based methods within a unified weighted sensitivity metric. The practical value of this approach is that its application in sensitivity analysis or neural network weight initialization improves the accuracy of parameter impact assessment, enhances model convergence, and reduces training time.