Skip to content
Tutorials Center
My learning

Topics

Lesson 4 of 41 ยท Getting Started

MLOps Roles and the Tool Landscape

Meet the people in an ML team and the families of MLOps tools (tracking, versioning, orchestration, serving, monitoring), and learn how to pick tools without chasing hype.

MLOps crosses several jobs. Data scientists, data engineers, ML engineers and platform engineers each hold a piece, and many small teams have one person wearing all four hats. Knowing the responsibilities makes the tools easier to understand, because every tool exists to serve one of them.

Who does what

Role Main concern Typical output
Data scientist Finding a model that solves the problem Experiments, evaluation reports
Data engineer Reliable, timely, clean data Data pipelines, tables, feature jobs
ML engineer Turning experiments into production services Training pipelines, packaged models, APIs
Platform / DevOps engineer Infrastructure, security, cost Clusters, CI/CD, observability
Product owner / domain expert Is the model useful and safe? Requirements, acceptance criteria

The handoffs between these roles are where most problems appear. MLOps tools and conventions make each handoff explicit: a registered model, a validated dataset, a versioned config.

Families of tools

No single product does everything. Think in families, then pick a tool per family.

Tool families

Data versioning DVC, lakeFS, Delta Lake.
Experiment tracking MLflow, Weights and Biases.
Orchestration Airflow, Kubeflow, Prefect.
Registry and serving MLflow, BentoML, KServe.

Pick per family, not one giant suite.

1 / 2

Product names change quickly, but the families are stable. The rest of this course teaches the concept in each family, builds a runnable stand-in, and shows the real tool's commands read-only.

Seeing the landscape as a pipeline

flowchart LR
  D["Data sources"] --> V["Validation and versioning"]
  V --> F["Features"]
  F --> T["Training and tracking"]
  T --> R["Registry"]
  R --> S["Serving"]
  S --> M["Monitoring"]
  M --> D

Each box has a family of tools and an owner.

Build, buy, or assemble

Three common strategies:

  1. Managed platform (SageMaker, Vertex AI, Azure ML, Databricks): fast start, integrated, but tied to a vendor and billed by usage.
  2. Open-source assembly (MLflow + Airflow + Docker + Prometheus): flexible and portable, but you operate it.
  3. Minimal homemade: scripts, Git, a cron job and a folder of artifacts. Cheapest, and perfectly fine for small systems.

Whichever you pick, keep model code independent of the platform so you can switch later.

A tool-selection helper

This runnable stand-in ranks strategies for your situation using simple weights. It is deliberately crude: the point is to make your assumptions explicit.

Python Pick a strategy

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

Change the numbers and watch the suggestion move. A larger team with strict regulation drifts toward managed or assembled stacks; a tiny team with one model stays minimal.

A sample tool inventory as data

Writing down your stack makes gaps visible. This example stores an inventory and reports which lifecycle stages have no tool.

Python Find the gaps in your stack

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

A gap is not automatically a problem. It is a prompt to ask whether the risk justifies adding a tool.

Avoid choosing a tool because it is popular. Pick the simplest option that solves the problem you actually have this quarter.

Check your understanding

Quick check ยท 4 questions

of 4 answered You scored / 4 ยท perfect!

  1. 1. Which role is mainly responsible for reliable, timely, clean data pipelines?

    Data engineers build and operate the data pipelines that feed ML systems.
  2. 2. Which tool family does MLflow Tracking belong to?

    MLflow Tracking records parameters, metrics and artifacts of runs.
  3. 3. What is a good reason to start with minimal scripts instead of a full platform?

    Match tooling to the real need; heavy platforms add operating cost.
  4. 4. Why keep model code independent of the platform?

    Portability limits vendor lock-in.

Summary

  • MLOps involves data scientists, data engineers, ML engineers and platform engineers.
  • Tools come in families: data versioning, tracking, orchestration, registry/serving, features, monitoring, CI/CD.
  • Strategies: managed platform, open-source assembly, or minimal scripts.
  • Pick tools from real needs and keep your code portable.

Next: The ML Lifecycle