On the Task of Designing Efficient Machine Learning Operations Architectures
Mikhail Gorodnichev
Abstract
This article is devoted to the development of support of information processes in the methodology of machine learning operations. The paper considers methods of optimization of information processes arising in the construction of modern intelligent transport systems, namely the deployment and operation of machine learning methods aimed at pattern recognition, anomaly detection, analysis of transactional data on the state of the traffic flow and prediction of loads on networks. The paper considers two types of information processing: batch and stream. Based on the subject area, the main focus is on stream data processing architectures and approaches. Taking into account the data format and information lifetime constraints, the following stream data processing architectures are proposed using: DBMS, GPU, Hadoop, S3 protocol, In Memory paradigm, HTTPS protocol, Kafka message broker. The proposed architectures are built on open-source tools, and allow for easy deployment of models. The developed architectures allow building fault-tolerant and scalable applications, reducing the cost of system operation.