Data & AI
MLOps
MLOps applies DevOps discipline to machine learning, and adds the problems DevOps never had: data that changes, models that degrade silently, and experiments that must be reproducible months later. It is the reason most models never reach production, and the fix.
Start here
You need one of the two halves already: either DevOps experience or ML experience. Starting with neither means learning two disciplines at once, which is slower than doing them in sequence.
Before you begin
- Comfortable with Docker and CI/CD, or with training models
- Python
- Basic cloud infrastructure
The two foundations
MLOps is the overlap. You need enough of both sides to be credible in either room.
RequiredThe engineering half
Model deployment is deployment. The general practices come first and transfer directly.
The engineering half
Model deployment is deployment. The general practices come first and transfer directly.
What to learn
- Containers
- CI/CD pipelines
- Infrastructure as code
- Cloud fundamentals
- Monitoring and alerting
Tools
- Docker
- GitHub Actions
- Terraform
- Kubernetes
Next: The ML half.
RequiredThe machine learning half
You cannot build a platform for a workflow you have never performed.
The machine learning half
You cannot build a platform for a workflow you have never performed.
What to learn
- Training a model end to end
- Evaluation and validation
- Feature engineering
- Why models degrade
- Reading someone else's training code
Tools
- Python
- scikit-learn or PyTorch
Practice
Next: Reproducibility.
Reproducibility
The first real MLOps problem: being able to recreate a result exactly, months later, on a different machine.
RequiredVersioning everything
A model is a function of code, data, parameters and environment. Versioning only the code reproduces nothing.
Versioning everything
A model is a function of code, data, parameters and environment. Versioning only the code reproduces nothing.
What to learn
- Code versioning
- Data versioning
- Environment pinning
- Experiment tracking
- Model registries with lineage
- Deterministic training where possible
Tools
- Git
- DVC
- MLflow
Practice
Project
intermediateA reproducible training setup
A repository where a single command reproduces a model and its metrics from raw data, with data, parameters and environment all versioned.
- DVC
- MLflow
- Docker
Someone else can clone your repository and get your numbers.
Next: Automating the pipeline.
Training pipelines
Turning a training script into a scheduled, monitored, restartable system.
RequiredAutomated training
Manual retraining does not happen. Whatever is not automated eventually stops being done.
Automated training
Manual retraining does not happen. Whatever is not automated eventually stops being done.
What to learn
- Pipeline orchestration
- Data validation before training
- Automated evaluation gates
- Conditional promotion to the registry
- Resource scheduling for GPUs
- Pipeline caching
Tools
- Kubeflow Pipelines
- Metaflow
- Airflow
- SageMaker Pipelines
Next: Features.
RecommendedFeature management
Solves training/serving skew and feature reuse. Also frequently adopted before the organisation has the problem it solves.
Feature management
Solves training/serving skew and feature reuse. Also frequently adopted before the organisation has the problem it solves.
What to learn
- Feature definitions as code
- Offline and online consistency
- Point-in-time correctness
- Feature reuse across teams
- When a feature store is overkill
Tools
- Feast
Next: Deployment.
Model deployment
Where MLOps looks most like DevOps, with one important difference: you can deploy a working model that is wrong.
RequiredServing patterns
How a model is served determines its latency, its cost and how safely it can be changed.
Serving patterns
How a model is served determines its latency, its cost and how safely it can be changed.
What to learn
- Batch and real-time inference
- Model servers
- Versioned endpoints
- Shadow deployment
- Canary and A/B testing models against each other
- Rollback
Tools
- KServe
- Seldon
- BentoML
- Managed endpoints
Practice
Next: Monitoring.
Monitoring and the feedback loop
The genuinely ML-specific part, and where the discipline earns its name. A model fails silently.
RequiredDrift and performance monitoring
An endpoint returning 200 with steadily worse predictions is the characteristic ML failure. Ordinary monitoring cannot see it.
Drift and performance monitoring
An endpoint returning 200 with steadily worse predictions is the characteristic ML failure. Ordinary monitoring cannot see it.
What to learn
- Data drift detection
- Concept drift
- Prediction distribution monitoring
- Delayed ground truth
- Segment-level performance
- Alerting on model quality
Tools
- Evidently
- A monitoring platform
Next: Closing the loop.
RequiredRetraining and the feedback loop
The full lifecycle: production data becomes training data, safely and without feedback loops that amplify the model's own bias.
Retraining and the feedback loop
The full lifecycle: production data becomes training data, safely and without feedback loops that amplify the model's own bias.
What to learn
- Retraining triggers
- Label collection
- Guarding against feedback loops
- Automated promotion with evaluation gates
- Keeping a human decision point
Next: Governance.
RecommendedGovernance and cost
Models make decisions about people and consume expensive compute. Both attract scrutiny.
Governance and cost
Models make decisions about people and consume expensive compute. Both attract scrutiny.
What to learn
- Model cards and documentation
- Audit trails from prediction to training data
- Approval workflows
- GPU cost attribution
- Access control on models and data
Project
advancedAn end-to-end ML platform
A system where a commit triggers validated training, evaluation against a baseline, registry promotion on pass, canary deployment, and drift monitoring with alerts that trigger retraining.
- MLflow
- A pipeline orchestrator
- Kubernetes
- A drift monitor
A model can go from commit to monitored production without a human copying a file, and you find out before users do when it degrades.
Where this leads
You do not have to pick one now. These are the directions this path opens up once you are working.
Continue your journey
The paths closest to this one. Skills overlap more than the job titles suggest.
Machine Learning Engineering
Build models that run in production, not only in notebooks.
ViewDevOps Engineering
Automate the path from a commit to something running in production.
ViewData Engineering
Build the pipelines that make everyone else's data work possible.
ViewPlatform Engineering
Build the internal product that lets every other engineer ship safely.
ViewYou do not have to do this alone
Our programs are free, taught live, and built around the same progression. Join one and work through it alongside other people.