Skip to main content
Containerization packages your pipeline code, dependencies, and environment into Docker images for reproducible execution across different infrastructure. ZenML handles Docker image building automatically while giving you full control over the build process.

​
Why Containerization Matters

Containerized orchestrators (Kubernetes, AWS SageMaker, GCP Vertex AI, etc.) run pipeline steps in isolated Docker containers. This ensures:

Reproducibility

Same environment across development, staging, and production

Isolation

Steps run in clean environments without conflicting dependencies

Portability

Run anywhere Docker runs - local, cloud, or on-premise

Version Control

Track exact versions of all dependencies and code

​
Docker Settings

Configure Docker builds using DockerSettings: 
from zenml import pipeline
from zenml.config import DockerSettings

docker_settings = DockerSettings(
    # Base image
    parent_image="python:3.9-slim",
    
    # Python packages
    requirements=["scikit-learn==1.3.0", "pandas==2.0.0"],
    
    # System packages
    apt_packages=["git", "wget"],
    
    # Environment variables
    environment={
        "LOG_LEVEL": "INFO",
        "CUSTOM_CONFIG": "production",
    },
)

@pipeline(settings={"docker": docker_settings})
def ml_pipeline():
    # Pipeline steps here
    pass

​
The ContainerizedOrchestrator Base Class

Orchestrators that run steps in containers inherit from ContainerizedOrchestrator: 
from zenml.orchestrators import ContainerizedOrchestrator
from typing import List
from zenml.config.build_configuration import BuildConfiguration

class MyOrchestrator(ContainerizedOrchestrator):
    """Orchestrator that runs steps in Docker containers."""
    
    @staticmethod
    def get_image(
        snapshot: PipelineSnapshotResponse,
        step_name: Optional[str] = None,
    ) -> str:
        """Get the Docker image for a pipeline or step.
        
        Args:
            snapshot: Pipeline snapshot
            step_name: Optional step name for step-specific image
            
        Returns:
            Image name or digest
        """
        if not snapshot.build:
            raise RuntimeError(f"Missing build for snapshot {snapshot.id}")
        
        return snapshot.build.get_image(
            component_key=ORCHESTRATOR_DOCKER_IMAGE_KEY,
            step=step_name,
        )
    
    def get_docker_builds(
        self, snapshot: PipelineSnapshotBase
    ) -> List[BuildConfiguration]:
        """Get required Docker builds.
        
        Args:
            snapshot: Pipeline snapshot
            
        Returns:
            List of build configurations
        """
        pipeline_settings = snapshot.pipeline_configuration.docker_settings
        builds = []
        
        # Check each step for custom Docker settings
        for name, step in snapshot.step_configurations.items():
            step_settings = step.config.docker_settings
            
            if step_settings != pipeline_settings:
                # Step has custom settings, needs separate image
                builds.append(
                    BuildConfiguration(
                        key=ORCHESTRATOR_DOCKER_IMAGE_KEY,
                        settings=step_settings,
                        step_name=name,
                    )
                )
            else:
                # Use pipeline-level image
                if not any(b.step_name is None for b in builds):
                    builds.append(
                        BuildConfiguration(
                            key=ORCHESTRATOR_DOCKER_IMAGE_KEY,
                            settings=pipeline_settings,
                        )
                    )
        
        return builds

​
Image Building Process

ZenML’s image building follows this process: 
┌─────────────────────────────────────────────────────────────┐
│ 1. Determine Base Image                                     │
│    - Custom dockerfile specified? Use it as base            │
│    - Otherwise: Use parent_image or ZenML default          │
└─────────────────────────────────────────────────────────────┘
                            │
                            ▼
┌─────────────────────────────────────────────────────────────┐
│ 2. Collect Requirements                                     │
│    - Local environment (if replicate_local_python_environment)│
│    - Stack requirements (if install_stack_requirements)     │
│    - Integration requirements (required_integrations)       │
│    - Pyproject.toml (if pyproject_path specified)          │
│    - Explicit requirements (requirements)                   │
└─────────────────────────────────────────────────────────────┘
                            │
                            ▼
┌─────────────────────────────────────────────────────────────┐
│ 3. Build Image                                              │
│    - Install system packages (apt_packages)                 │
│    - Set environment variables (environment)                │
│    - Install Python packages                                │
│    - Copy code files (if allowed)                           │
│    - Set runtime environment (runtime_environment)          │
└─────────────────────────────────────────────────────────────┘
                            │
                            ▼
┌─────────────────────────────────────────────────────────────┐
│ 4. Push to Registry                                         │
│    - Tag with target_repository (or default)                │
│    - Push to container registry from stack                  │
└─────────────────────────────────────────────────────────────┘

​
Requirements Installation Order

Dependencies are installed in this specific order: 
# 1. Local environment packages (optional)
if docker_settings.replicate_local_python_environment:
    # Captures pip freeze or uv freeze output
    install_from_local_environment()

# 2. Stack requirements (optional)
if docker_settings.install_stack_requirements:
    # Installs requirements for all stack components
    install_stack_dependencies()

# 3. Integration requirements
for integration in docker_settings.required_integrations:
    # Installs ZenML integration packages
    install_integration(integration)

# 4. Pyproject.toml (optional)
if docker_settings.pyproject_path:
    # Exports and installs from pyproject.toml
    install_from_pyproject()

# 5. Explicit requirements
if docker_settings.requirements:
    # Installs from requirements file or list
    install_requirements()

​
Common Configuration Patterns

​
Basic Configuration

Simple setup with explicit requirements: 
from zenml.config import DockerSettings

docker_settings = DockerSettings(
    requirements=[
        "scikit-learn==1.3.0",
        "pandas==2.0.0",
        "numpy==1.24.0",
    ],
)

​
Using Requirements File

# requirements.txt in your project root
# scikit-learn==1.3.0
# pandas==2.0.0
# numpy==1.24.0

docker_settings = DockerSettings(
    requirements="requirements.txt",
)

​
Custom Base Image

docker_settings = DockerSettings(
    parent_image="pytorch/pytorch:2.0.0-cuda11.7-cudnn8-runtime",
    requirements=["transformers==4.30.0"],
)

​
Custom Dockerfile

For complete control: 
# Dockerfile
FROM nvidia/cuda:11.8.0-cudnn8-runtime-ubuntu22.04

# Install Python
RUN apt-get update && apt-get install -y python3.9 python3-pip

# Install ZenML
RUN pip install zenml==0.55.0

# Install ML libraries
RUN pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

docker_settings = DockerSettings(
    dockerfile="Dockerfile",
    build_context_root=".",  # Directory containing Dockerfile
)

​
System Dependencies

docker_settings = DockerSettings(
    apt_packages=["git", "wget", "curl", "build-essential"],
    requirements=["gitpython"],
)

​
Environment Variables

docker_settings = DockerSettings(
    # Set during build (available during requirement installation)
    environment={
        "PIP_INDEX_URL": "https://custom.pypi.org/simple",
        "BUILD_CONFIG": "production",
    },
    
    # Set at runtime (available during step execution)
    runtime_environment={
        "LOG_LEVEL": "INFO",
        "API_KEY": "${SECRET_API_KEY}",  # Reference from secrets
    },
)

​
Integration Requirements

docker_settings = DockerSettings(
    # Automatically install integration dependencies
    required_integrations=["sklearn", "mlflow", "s3"],
    
    # Stack requirements auto-detected
    install_stack_requirements=True,
)

​
Step-Specific Docker Settings

Different steps can use different images: 
from zenml import pipeline, step
from zenml.config import DockerSettings

# Default settings for most steps
pipeline_docker = DockerSettings(
    requirements=["pandas==2.0.0"],
)

# Special settings for GPU training
training_docker = DockerSettings(
    parent_image="pytorch/pytorch:2.0.0-cuda11.7-cudnn8-runtime",
    requirements=["transformers==4.30.0"],
)

@step
def load_data() -> dict:
    # Uses pipeline_docker
    return {"data": "loaded"}

@step(settings={"docker": training_docker})
def train_model(data: dict) -> dict:
    # Uses training_docker with GPU support
    import torch
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    return {"model": "trained", "device": str(device)}

@pipeline(settings={"docker": pipeline_docker})
def ml_pipeline():
    data = load_data()
    model = train_model(data)

​
Dynamic Pipelines and Docker

For dynamic pipelines, the orchestration container needs its own image: 
class ContainerizedOrchestrator(BaseOrchestrator):
    def should_build_pipeline_image(self, snapshot: PipelineSnapshotBase) -> bool:
        """Whether to build the pipeline image.
        
        Args:
            snapshot: Pipeline snapshot
            
        Returns:
            True for dynamic pipelines (need orchestration container)
        """
        return snapshot.is_dynamic
The orchestration container:
  • Runs the pipeline function to discover steps
  • Submits discovered steps to the orchestration backend
  • Typically needs modest resources (CPU and memory only)

​
Advanced Build Options

​
Build Options

Customize Docker build behavior: 
from zenml.config import DockerSettings, DockerBuildConfig, DockerBuildOptions

docker_settings = DockerSettings(
    build_config=DockerBuildConfig(
        build_options=DockerBuildOptions(
            # Always pull base image
            pull=True,
            
            # Don't use cache
            no_cache=False,
            
            # Build arguments
            build_args={
                "PYTHON_VERSION": "3.9",
                "BUILD_ENV": "production",
            },
            
            # Cache from images
            cache_from=[
                "myregistry/ml-base:latest",
            ],
            
            # Labels
            labels={
                "version": "1.0.0",
                "team": "ml-platform",
            },
        ),
    ),
)

​
Package Installers

Choose between pip and uv: 
from zenml.config import DockerSettings, PythonPackageInstaller

# Use uv (faster, default)
docker_settings = DockerSettings(
    python_package_installer=PythonPackageInstaller.UV,
    requirements=["pandas", "scikit-learn"],
)

# Use pip
docker_settings = DockerSettings(
    python_package_installer=PythonPackageInstaller.PIP,
    python_package_installer_args={
        "index_url": "https://custom.pypi.org/simple",
    },
    requirements=["pandas", "scikit-learn"],
)

​
Code Inclusion

Control how code is made available: 
docker_settings = DockerSettings(
    # Allow including files in image
    allow_including_files_in_images=True,
    
    # Allow downloading from code repository
    allow_download_from_code_repository=True,
    
    # Allow downloading from artifact store
    allow_download_from_artifact_store=True,
)
Code resolution priority:
  1. Download from code repository (if configured and allowed)
  2. Download from artifact store (if allowed)
  3. Include in Docker image (if allowed)

​
Running as Non-Root User

docker_settings = DockerSettings(
    # Run container as specific user
    user="zenml",
    
    # User becomes owner of /app directory
    requirements=["pandas"],
)

​
Testing Docker Builds Locally

Test your Docker configuration before running on cloud infrastructure: 
# Use the local Docker orchestrator
from zenml import pipeline

@pipeline(name="test_pipeline")
def test_docker_config():
    # Your pipeline steps
    pass

# Run with local Docker orchestrator
test_docker_config.with_options(
    orchestrator="local_docker",
)()
This builds the Docker image locally and runs steps in containers on your machine.

​
Best Practices

Pin Versions

Always specify exact versions for reproducibility: pandas==2.0.0 not pandas

Minimize Image Size

Use slim base images and only install necessary packages

Layer Caching

Order Dockerfile commands from least to most frequently changing

Security Scanning

Regularly update base images and scan for vulnerabilities

​
Security Considerations

# DON'T: Hardcode secrets in Docker settings
docker_settings = DockerSettings(
    environment={
        "API_KEY": "secret-key-12345",  # ❌ Bad!
    }
)

# DO: Reference secrets from ZenML secret store
docker_settings = DockerSettings(
    runtime_environment={
        "API_KEY": "${SECRET_API_KEY}",  # ✅ Good!
    }
)

​
Multi-Stage Builds

For complex build requirements: 
# Dockerfile with multi-stage build
FROM python:3.9 AS builder

WORKDIR /build
COPY requirements.txt .
RUN pip install --user -r requirements.txt

FROM python:3.9-slim

# Copy installed packages from builder
COPY --from=builder /root/.local /root/.local

# Install ZenML
RUN pip install zenml==0.55.0

# Update PATH
ENV PATH=/root/.local/bin:$PATH

​
Next Steps

Custom Orchestrators

Build orchestrators that work with Docker images

Resource Configuration

Configure resources for containerized steps

Dynamic Pipelines

Understand orchestration containers for dynamic pipelines

Custom Materializers

Handle data between containerized steps

Build docs developers (and LLMs) love

Get started for freeTalk to us