2 What is MLOps ?

MLOps is presented as the set of practices that transform machine learning from isolated modeling into a repeatable, production-grade capability that reliably delivers business value. Because models, data, and assumptions evolve, the chapter frames ML as a closed, iterative loop rather than a one-off build: start with a well-aligned problem definition and success metrics, then continuously learn from outcomes to refine the system. The core idea is to bridge the gap between business goals, technical requirements, and operational constraints through shared processes, clear ownership, and rigorous tracking so that models can be changed quickly and safely without sacrificing confidence.

The chapter walks through each stage of the loop—data collection, EDA, modeling and training, evaluation, deployment, monitoring, and maintenance—emphasizing lineage, versioning, and automation. Data must be relevant, representative, and carefully tracked; EDA validates assumptions and informs feature choices; modeling benefits from modular code, experiment tracking, and hyperparameter search to maximize reproducibility and velocity. Evaluation uses appropriate domain metrics and robust holdouts, including error analysis and (optionally) interpretability techniques. Deployment spans APIs and edge targets with environment-specific optimizations and staged rollouts. Monitoring detects drift, performance regressions, and errors, backed by alerting and strong logging. Maintenance closes the loop by feeding insights back into data, models, and infrastructure for continuous improvement.

Robust MLOps is necessary because real-world ML adds complexities that differ from traditional software: data is a first-class asset, models change without code edits, and compliance, bias, and drift must be actively managed. The chapter contrasts DevOps and MLOps—sharing automation and CI/CD principles but diverging on data stewardship, continuous training, interpretability, and performance monitoring. It also outlines organizational challenges (tooling fragmentation, cross-functional communication, scaling/optimization) and the benefits of maturity: faster experimentation, cost control, collaboration, repeatability, traceability, and reliable scaling. A maturity model (Level 0: manual, Level 1: continuous retraining pipelines, Level 2: pipeline automation) provides a path forward, underscoring that disciplined, automated processes reduce technical debt and build lasting confidence in production ML.

The mental map where we are focusing on defining the problem(1) and model design(2)

ML as a loop

Examples of the visual data in the MIDV500 dataset

Example of an annotated ID card, shown in CVAT, which is a web-based tool designed for annotating images and videos, commonly used to label data for computer vision models.

A view of a retraining pipeline using the modular codebase concept. This approach of keeping the model, code, configuration files, and data as distinct versioned components with lineage links ensures that the process remains flexible, fast, and adaptable while enabling experimentation, debugging, and iterative development.

• ML exists to solve a business problem and it is important to understand the requirement in depth before starting an ML project.
• MLOps is the iterative process of developing, monitoring and improving an ML model.
• A model is an artifact of the ML loop that aims to improve model performance over time.
• MLOps is hard due to data management, complex tooling, organizational setups, scaling challenges and the unpredictability of the real world.
• Skipping established ML practices can appear to be faster in the short term, but duplication and technical debt will quickly erase any gains.
• DevOps and MLOps have similarities, but differences in data and model management, among others, means that MLOps has some unique challenges.
• Robust MLOps is a highly experimental, iterative process with room for institutional learning and rapid prototyping to identify things that work for you and your organization.

FAQ