ZenML: MLOps for Reliable AI

0.94.3 verified Fri May 01 auth: no python

ZenML is an open-source MLOps framework that helps you build reproducible, production-ready ML pipelines. It provides a unified interface for orchestrating pipelines, managing infrastructure, and tracking experiments across various backends (local, cloud, Kubernetes). The current version is 0.94.3, with frequent releases (multiple per month). It requires Python >=3.10,<3.14.

pip install zenml

Common errors

error ModuleNotFoundError: No module named 'zenml.core' ↓

cause Using old import path that was removed in v0.20+.

fix

Use from zenml import pipeline, step instead.

error ImportError: cannot import name 'ArtifactConfig' from 'zenml.artifacts' ↓

cause ArtifactConfig moved to the top-level module in v0.93+.

fix

Use from zenml import ArtifactConfig.

error TypeError: 'ArtifactVersionResponse' object is not iterable ↓

cause Code expects a single artifact version, but in v0.93+ `regular_inputs` returns a list for each key.

fix

Iterate over the list: for artifact in step_run.regular_inputs['key']: ...

error sqlalchemy.exc.OperationalError: (sqlite3.OperationalError) no such table: ... ↓

cause ZenML server database schema not migrated; older version of database used with newer server.

fix

Run zenml upgrade or manually delete and reinitialize the database (careful with production data).

Warnings

breaking In v0.93.0, `StepRunResponse.regular_inputs` changed from `Dict[str, ArtifactVersionResponse]` to `Dict[str, List[ArtifactVersionResponse]]`. Code relying on old type will break. ↓

fix Update code to handle lists of artifact versions instead of single artifacts.

gotcha The `zenml init` command is not needed unless you're using ZenML Pro or legacy project structure. For most users, simply installing and importing ZenML is sufficient. ↓

fix Skip `zenml init` unless instructed by the documentation for ZenML Pro.

breaking In v0.93.0, database migration may fail if pipelines without any runs exist. Delete such pipelines before upgrading server. ↓

fix Run the provided script to delete pipelines with `latest_run_status is None` before server upgrade.

gotcha Custom step operator flavors must implement new `submit_step` and `get_step_status` methods (v0.94+). Legacy `launch` method only works for static pipelines. ↓

fix Update custom step operators to implement the new methods.

Install

pip install zenml[server]

pip install zenml[templates,server]

Imports

pipeline
wrong
```
from zenml.core import pipeline
```
correct
```
from zenml import pipeline
```
Incorrect import path; causes ModuleNotFoundError.
step
wrong
```
from zenml.steps import step
```
correct
```
from zenml import step
```
Deprecated in v0.20+; use central imports.

Client

wrong

from zenml import Client

correct

from zenml.client import Client

Client was moved to zenml.client in v0.50+.

ArtifactConfig

wrong

from zenml.artifacts import ArtifactConfig

correct

from zenml import ArtifactConfig

ArtifactConfig moved to top-level in v0.93+.

Quickstart

Define and run a simple two-step pipeline locally.

from zenml import pipeline, step

@step
def load_data() -> dict:
    return {'data': [1, 2, 3]}

@step
def train_model(data: dict) -> str:
    return 'model trained with ' + str(len(data['data'])) + ' samples'

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

if __name__ == '__main__':
    my_pipeline()
    print('Pipeline run completed')