The Components of an MLOps System
"MLOps" can sound like one tool you install. It is really a set of parts that hand work to each other in a loop. Once you can name each part and what it produces, every tool in the rest of this course has an obvious place to live.
The rule
Data in, model out, watch it, loop back.
Follow the Tunebox churn model around the loop. Each component takes something in and produces something the next one needs:
| Component | Produces | Example tool |
|---|---|---|
| Version control | a pinned version of code and data | Git, DVC |
| Training pipeline | a trained model, run in fixed steps | Airflow, DVC pipelines |
| Experiment tracking | a record of every run's settings and scores | MLflow |
| Model registry | named, numbered model versions and which is live | MLflow registry |
| CI/CD | automatic checks and releases for each change | GitHub Actions |
| Serving | predictions on request or in batches | a REST API in Docker |
| Monitoring | signals that the model or data has changed | Prometheus, Grafana |
Two more ideas thread through all of them. Orchestration is what runs the steps in the right order at the right time. Data lineage is the chain of records that says which data and code produced which model, and which model produced which prediction.
The last arrow matters most: monitoring feeds back into training. When live data drifts away from what the model learned, the loop runs again. That is why MLOps is drawn as a circle, not a line.
The shortcut
For any tool or task, ask "What does it take in, and what does it hand on?" The answer places it in the loop.
Version
A new month of subscription data is committed with DVC; Git holds the code that will use it.
Drills
Answer fast — trust the rule.
Which component answers "which model version is live right now?"
Where do you look to compare the recall of 40 training runs with different settings?
What turns the MLOps diagram from a line into a loop?
A prediction looks wrong and you need to know which data trained the model that made it. Which idea are you relying on?
Remember
- Data in, model out, watch it, loop back.
- Place any tool by asking what it takes in and what it hands on.
- Trap: thinking MLOps is one product; it is a loop of cooperating parts.