🎯 Introduction

Deploying a model to production is not the finish line—it's the starting line. Machine learning systems require continuous attention: models drift as the world changes, pipelines break under new data conditions, and teams struggle to reproduce results from months ago. MLOps brings software engineering rigor to these operational challenges.

This page covers design patterns that address the full lifecycle of ML systems in production:

• Workflow orchestration: Structuring complex ML pipelines as reproducible, containerized DAGs
• Version management: Deploying and deprecating model versions without breaking downstream systems
• Continuous evaluation: Detecting model degradation before it impacts business outcomes

These patterns bridge the gap between experimental notebooks and reliable production systems—turning one-off model training into sustainable, maintainable ML infrastructure.

⚙️ Workflow Pipeline

Individual notebooks don't scale to production. Workflow pipelines organize ML processes as containerized steps in a directed acyclic graph (DAG).

⚠️ Problem

Individual data scientists can run preprocessing, training, and deployment steps from a single script or notebook. However, this approach doesn't scale as:

• ML workflows become more complex with multi-model systems
• More team members need to contribute and collaborate
• Reproducibility requirements increase (regulatory compliance, audit trails)
• Different steps require different dependencies and compute resources
• Production deployment demands reliability and automation

✅ Solution

Organize the ML workflow as a directed acyclic graph (DAG) where each step runs in its own container.

<aside> 💡

Open-source Orchestrators

• Kubeflow Pipelines: Kubernetes-native, framework-agnostic orchestration with strong ML focus
• Flyte: Kubernetes-native platform built for reproducibility with strong typing and versioning
• Apache Airflow: Industry-standard DAG-based orchestration with extensive integrations
• Metaflow: Netflix-originated tool designed for data scientists to focus on models over infrastructure </aside>