What Is MLOps? The Machine Learning Lifecycle in Production

MLOps (machine learning operations) is the set of practices for getting machine learning models into production and keeping them useful. A model in a notebook is a one-off result. A model in production must be reproducible, deployed safely, fed fresh data, monitored and updated as the world changes. MLOps borrows ideas from DevOps (automation, testing, monitoring) and data engineering (reliable pipelines), and adds problems specific to models.

Why models are harder to run than normal software

• Behaviour comes from data, not only code. The same code trained on different data gives a different model.
• Models decay. Customer behaviour, prices and fraud tactics change, so a model that was accurate last year slowly gets worse.
• Failures are quiet. A broken model often keeps returning plausible-looking predictions instead of crashing.
• Training and serving can drift apart. A feature calculated one way in training and another way in production silently breaks predictions.

The lifecycle

1. Data: collect, validate and version the training data.
2. Experiment: try features, models and settings, and track every run.
3. Train: build the final model with a reproducible pipeline.
4. Validate: compare with the current model on held-out data and agreed thresholds.
5. Register: store the model with its metadata, so it can be found, compared and rolled back. See experiment tracking and model registry.
6. Deploy: serve predictions in batch or online. See model serving.
7. Monitor: watch inputs, predictions, outcomes and system health. See monitoring and drift.
8. Retrain: on a schedule or when monitoring shows the model has degraded, then repeat.

What to version

To reproduce any model in production, you need:

Item	Why
Code	The training and feature logic
Data	Exactly which rows trained it (a snapshot, or a fingerprint and a pointer)
Settings	Hyperparameters and random seeds
Environment	Library versions, ideally a container image
Model file and metrics	What was deployed and how good it was

A simple first setup

You do not need a big platform for a first model:

• A training script in Git, run by a scheduler, that writes the model and a metadata file.
• A folder or bucket per model version (a basic registry), with a simple promotion rule.
• A nightly batch job that writes predictions to a table.
• A dashboard of prediction volumes, input distributions and, when available, real outcomes.

Add dedicated tools for experiment tracking, feature stores and serving as the number of models and teams grows.

Common mistakes

1. Notebooks as production. Move training into tested, scheduled code.
2. No record of what trained a model, so nobody can reproduce or explain it.
3. No monitoring, so degraded models run for months.
4. Over-engineering early: a heavy platform for one model.

Key takeaway

MLOps keeps models reproducible, deployed safely and monitored, through a lifecycle of data, experiments, training, validation, registration, deployment, monitoring and retraining. Version code, data, settings and environment for every model, start simple, and add tools as needs grow. Next: experiment tracking and model registry.