MLFlow+: Towards Reliable and Debuggable MLOps with Unified Lineage and Built-in Observability

The operationalization of machine learning (ML) systems in production remains fraught with challenges, including fragmented tooling, poor reproducibility, and inadequate monitoring. While existing MLOps platforms such as MLflow, Kubeflow, and SageMaker Pipelines address isolated aspects of the ML lifecycle, they often lack native support for end-to-end lineage, real-time observability, and extensibility without custom integration layers. To bridge this gap, we introduce MLFlow+, an open-source, extensible MLOps framework that unifies experimentation, deployment, and monitoring into a single, coherent architecture. At its core, MLFlow+ employs a temporal metadata graph that captures causal relationships between data versions, code commits, model artifacts, and runtime telemetry. The framework natively incorporates statistical drift detectors (Kolmogorov-Smirnov for continuous features, Population Stability Index for categorical features) and provides a lightweight plugin SDK for custom observability logic, alerting, and compliance policies. We rigorously evaluate MLFlow+ against six state-of-the-art MLOps baselines using three real-world datasets spanning fintech, telecommunications, and healthcare. Our results demonstrate that MLFlow+ reduces mean time to diagnose (MTTD) by ۶۲.۳% (p <۰.۰۱), improves deployment success rate by ۱۷.۲ percentage points, and introduces less than ۴% CPU overhead during inference. An ablation study confirms that the integration of unified lineage and built-in observability are the primary drivers of these gains. MLFlow+ is publicly available under the Apache ۲.۰ license at: https://github.com/somayeh-karimi/mlflow-plus

کلیدواژه ها:

MLOps ، Machine Learning Lifecycle ، Model Monitoring ، Data Drift ، Lineage Tracking ، Extensible Framework ، Production ML ، Debuggability ، Observability