MLOps
The ML lifecycle in production: versioning, registries, CI/CD, infrastructure, automation, monitoring and drift.
- lessons
- 7
- steps
- 60
- time
- 86 min
MLOps and the ML Lifecycle
MLOps turns a one-off notebook model into a repeatable loop: collect, prepare, train, evaluate, deploy, monitor — and go round again.
- What MLOps adds on top of DevOps for ML systems
- The six lifecycle stages and why they form a loop
- What happens inside collection, preprocessing and training
- How evaluation gates deployment and how models are rolled out safely
Versioning Data, Code and Models
An ML model is the product of a dataset and a commit — version all three and link them, and you can reproduce, compare and roll back anything.
- Why ML needs three linked version histories, not just Git
- How DVC versions large datasets with hashes and pointer files
- How model versions enable comparison and rollback
- What MLflow records for every training run
The Model Registry
One governed home for every model: artifacts in storage, searchable metadata, lineage, stages and approval gates.
- What a model registry is and who reads and writes it
- How artifacts are stored and what metadata makes them findable
- How lineage records where a model came from
- How models move through development → staging → production
CI/CD for Machine Learning
Automate the path from a commit or a new batch of data to a tested, safely-deployed model — with automatic rollback when it goes wrong.
- What CI/CD for ML triggers on, beyond code commits
- The stages of an automated training pipeline
- The five kinds of ML tests and what each catches
- How a deployment pipeline rolls out and validates a model
ML Infrastructure and Orchestration
Pipelines as DAGs, the orchestrators that run them — Airflow, Kubeflow, TFX — and the CPUs, GPUs and TPUs underneath.
- The three layers of ML infrastructure
- What an ML pipeline manages beyond a script
- Choosing and managing CPUs, GPUs and TPUs
- How Airflow, Kubeflow and TFX differ
Monitoring, Drift and Degradation
Models fail silently: inputs shift, the world changes, accuracy slides. Measure against a baseline, detect drift, and alert someone who can act.
- Why a model can be broken while every system dashboard is green
- What performance monitoring tracks for a model
- How data drift is detected with statistical tests per feature
- How concept drift differs from data drift
Automating the ML Loop
Retraining, feature engineering, hyperparameter tuning, model selection and whole workflows — run by triggers instead of people, with guardrails.
- The three kinds of trigger that start automated work
- How champion/challenger retraining deploys only improvements
- Grid, random and Bayesian hyperparameter search
- Chaining everything into an event-driven workflow with error handling