Halyk Bank with the support of Databases has implemented an enterprise ModelOps platform based on open-source technologies

2022: Enterprise ModelOps Platform Implementation

Halyk Bank implemented a corporate ModelOps platform deployed on the basis ON of s. The author open source of the project was a team of experts from an international company. integrator Databorn The solution made it possible to reduce at least twice time the putting into commercial operation (time-to-market) - ML models. This was announced on January 12, 2023 TAdviser by Andrey Lyunts, director of Databases.

The need for a technological solution was due to the lack of a single platform in the bank for the development and use of ML models, as well as a unified system for tracking their metrics and parameters.

Over the past years, we have been actively developing the Data Science direction. The number of ML models is increasing, and it has become more difficult to do without a full-fledged production system for the models being developed, "explained Roman Mashchyk, Deputy Chairman of the Board of Halyk Bank. - To resolve the issue, we engaged the expert team Databases, which implemented the corporate ModelOps platform. Now all ML models are developed on the basis of a single template and have a standardized pipeline of productivity for routine prediction and automatic retraining.

The platform is integrated with banking data sources, consists of various tools deployed in the Kubernetes cluster. This solution allows for tool flexibility, efficient management of available computing resources, and rapid scalability.

CI/CD is used for continuous integration (packaging) and production Gitlab. With its help, a single pipeline of model output from the development stage to use in production was built. MLFlow is used as a tool for managing Data Science experiments, which allows you to log the metrics and parameters of the model, as well as various artifacts of experiments, make decisions on the implementation of models, and perform a retrospective analysis of the process of changing metrics. The orchestrator of the application of ML models is Airflow.

Based on the results of pilot operation, using the model transferred to the platform as an example, on average, the entire process from model creation to commissioning is now at least 2 times faster than before the platform was implemented.

MLOps helps the business develop the Data Science direction more efficiently and implement high-quality ML models much faster, "said Andrey Lyunts. - The approach combines Machine Learning, DevOps, Data Engineering and Model Governance into a single methodology for creating, implementing and operating machine learning models. CI/CD processes on the platform are built into clear and uniform steps: development, assembly, testing, implementation and operation, which allows you to produce the model faster. The acceleration of the time-to-market frees up time for the development of new models by the bank's specialists, which allows you to solve more business problems using machine learning technologies.

Halyk Bank aims to develop the Data Science direction intensively, and the introduction of ModelOps will play an important role in achieving this goal.