ZenML Tutorial: Build Your First Production-Ready ML Pipeline in Minutes

In this tutorial you will build your first ZenML pipeline — from installing the CLI to running a reproducible ML workflow with caching and versioning. ZenML is a Python-native framework, so everything happens in a few lines of code.

Step 1: Install and Connect

Install ZenML with pip, then point the client at your self-hosted server (or use the local default).

pip install zenml
zenml init
zenml connect --url http://localhost:8080

Step 2: Define Your First Pipeline

A pipeline is just a Python function decorated with @pipeline. Each step gets its own decorator and runs in isolation:

from zenml import pipeline, step

@step
def load_data() -> dict:
    return {"samples": 1000, "features": 8}

@step
def train_model(data: dict) -> str:
    return f"trained on {data['samples']} samples"

@pipeline
def training_pipeline():
    data = load_data()
    train_model(data)

if __name__ == "__main__":
    training_pipeline()

Step 3: Run It — Then Run It Again

Execute the script. The first run trains your model; the second run is cached — ZenML detects unchanged steps and skips them. This is where production MLOps starts: every run is versioned, every artifact is tracked, and nothing is silently lost.

Step 4: Visualize and Track

Open the ZenML dashboard to inspect pipeline runs, compare model versions, and browse artifacts. The dashboard is served by the ZenML server — the same container you deploy with docker-compose.

Next Steps: Add an Agent Workflow

Once your pipeline is solid, extend it to the agent era: add RAG retrieval steps, evaluate LLM responses, and orchestrate agent calls as first-class pipeline steps. ZenML gives you one framework for classical ML and LLMOps alike.