A Platform to Manage the End-to-End Lifecycle of Batch-Prediction Machine Learning Models

Abstract

Developing, deploying and monitoring in production setups are all aspects of the end-to-end lifecycle of machine intelligence workflows. Even though recognized to play key roles in this lifecycle, reproducibility and automation are often enough neglected, leading to undesired long-term results. In this paper we outline the typical reproducibility and automation pitfalls, and subsequently introduce the platform we built to address these during training, validation, productionization and operation. Tuned for batch-predictions on large volumes of data, the platform also focuses on zero-rework production deployment and monitoring. We also present the case study of an ensemble model developed and deployed on our platform for a multi-dimensional image analysis problem.