Trigger pipelines from external systems

A step-by-step tutorial on effectively triggering your ZenML pipelines from external systems

This tutorial demonstrates practical approaches to triggering ZenML pipelines from external systems. We'll explore multiple methods: ZenML Pro's Snapshots for batch-style triggering, Pipeline Deployments for persistent HTTP services (available in OSS and Pro), and a custom API approach for fully bespoke solutions.

Introduction: The Pipeline Triggering Challenge

In development environments, you typically run your ZenML pipelines directly from Python code. However, in production, pipelines often need to be triggered by external systems:

• Scheduled retraining of models based on a time interval

• Batch inference when new data arrives

• Event-driven ML workflows responding to data drift or performance degradation

• Integration with CI/CD pipelines and other automation systems

• Invocation from custom applications via API calls

Each scenario requires a reliable way to trigger the right version of your pipeline with the correct parameters, while maintaining security and operational standards.

For full reference documentation, see the Snapshot docs for batch-style triggering and the Pipeline Deployment docs for persistent HTTP services.

Prerequisites

Before starting this tutorial, make sure you have:

1. ZenML installed and configured

2. Basic understanding of ZenML pipelines and steps

3. A simple pipeline to use for triggering examples

Creating a Sample Pipeline for External Triggering

First, let's create a basic pipeline that we'll use throughout this tutorial. This pipeline takes a dataset URL and model type as inputs, then performs a simple training operation:

from typing import Dict, Any, Union
from zenml import pipeline, step
import numpy as np
import pandas as pd
from sklearn.ensemble import RandomForestClassifier, GradientBoostingClassifier
from sklearn.model_selection import train_test_split
from sklearn.metrics import accuracy_score

@step
def load_data(data_url: str) -> pd.DataFrame:
    """Load data from a URL (simulated for this example)."""
    # For demonstration, we'll create synthetic data
    np.random.seed(42)
    n_samples = 1000
    
    print(f"Loading data from: {data_url}")
    # In a real scenario, you'd load from data_url
    # E.g., pd.read_csv(data_url)
    
    data = pd.DataFrame({
        'feature_1': np.random.normal(0, 1, n_samples),
        'feature_2': np.random.normal(0, 1, n_samples),
        'feature_3': np.random.normal(0, 1, n_samples),
        'target': np.random.choice([0, 1], n_samples)
    })
    return data

@step
def preprocess(data: pd.DataFrame) -> Dict[str, Any]:
    """Split data into train and test sets."""
    X = data.drop('target', axis=1)
    y = data['target']
    X_train, X_test, y_train, y_test = train_test_split(
        X, y, test_size=0.2, random_state=42
    )
    return {
        'X_train': X_train, 
        'X_test': X_test, 
        'y_train': y_train, 
        'y_test': y_test
    }

@step
def train_model(
    datasets: Dict[str, Any], 
    model_type: str = "random_forest"
) -> Union[RandomForestClassifier, GradientBoostingClassifier]:
    """Train a model based on the specified type."""
    X_train = datasets['X_train']
    y_train = datasets['y_train']
    
    if model_type == "random_forest":
        model = RandomForestClassifier(n_estimators=100, random_state=42)
    elif model_type == "gradient_boosting":
        model = GradientBoostingClassifier(random_state=42)
    else:
        raise ValueError(f"Unknown model type: {model_type}")
    
    print(f"Training a {model_type} model...")
    model.fit(X_train, y_train)
    return model

@step
def evaluate(
    datasets: Dict[str, Any], 
    model: Union[RandomForestClassifier, GradientBoostingClassifier]
) -> Dict[str, float]:
    """Evaluate the model and return metrics."""
    X_test = datasets['X_test']
    y_test = datasets['y_test']
    
    y_pred = model.predict(X_test)
    accuracy = accuracy_score(y_test, y_pred)
    
    print(f"Model accuracy: {accuracy:.4f}")
    return {'accuracy': float(accuracy)}


@pipeline
def training_pipeline(
    data_url: str = "s3://example-bucket/data.csv",
    model_type: str = "random_forest"
):
    """A configurable training pipeline that can be triggered externally."""
    data = load_data(data_url)
    datasets = preprocess(data)
    model = train_model(datasets, model_type)
    metrics = evaluate(datasets, model)

# For local execution during development
if __name__ == "__main__":
    # Run with default parameters
    training_pipeline()

This pipeline is designed to be configurable with parameters that might change between runs:

• data_url: Where to find the input data

• model_type: Which algorithm to use

These parameters make it an ideal candidate for external triggering scenarios where we want to run the same pipeline with different configurations.

Method 1: Using Snapshots (ZenML Pro)

This is a ZenML Pro-only feature. Please sign up here to get access.

Important: Workspace API vs ZenML Pro API

Snapshots use your Workspace API (your individual workspace URL), not the ZenML Pro API (cloudapi.zenml.io). This distinction is crucial for authentication - you'll need to use ZenML Pro credentials with the Workspace API, not the ZenML Pro management API. See ZenML Pro Personal Access Tokens and ZenML Pro Organization Service Accounts.

Production authentication (ZenML Pro)

For production automation in Pro (running snapshots from CI/CD or external systems), you can use Personal Access Tokens or Organization Service Accounts. Set ZENML_STORE_URL to your workspace URL and ZENML_STORE_API_KEY to your Personal Access Token or Organization Service Account API key.

Snapshots are the most straightforward way to trigger pipelines externally in ZenML. They provide a pre-defined, parameterized configuration that can be executed via multiple interfaces.

Creating a Snapshot

First, we need to create a snapshot of our pipeline. This requires having a remote stack with at least a remote orchestrator, artifact store, and container registry.

# The source path is the module path to your pipeline
zenml pipeline snapshot create <PIPELINE_SOURCE_PATH> \
    --name=production-training-template

You can also pass a config file and specify a stack:

# Create a config file
echo "steps:
  load_data:
    parameters:
      data_url: s3://production-bucket/latest-data.csv" > config.yaml

zenml pipeline snapshot create <PIPELINE_SOURCE_PATH> \
    --name=<TEMPLATE_NAME> \
    --config=<PATH_TO_CONFIG_YAML> \
    --stack=<STACK_ID_OR_NAME>

Running a snapshot

Once you have created a snapshot, there are multiple ways to run it, either programmatically with the Python client or via REST API for external systems.

Using the Python Client:

from zenml.client import Client

# Find snapshots for a specific pipeline
snapshots = Client().list_snapshots(pipeline=<PIPELINE-NAME-OR-ID>)

if snapshots:
    snapshot = snapshots[0]
    print(f"Using snapshot: {snapshot.name} (ID: {snapshot.id})")
    
    config = snapshot.config_template
    
    # Update the configuration with step parameters
    # Note: Parameters must be set at the step level rather than pipeline level
    config["steps"] = {
        "load_data": {
            "parameters": {
                "data_url": "s3://test-bucket/latest-data.csv",
            }
        },
        "train_model": {
            "parameters": {
                "model_type": "gradient_boosting",
            }
        }
    }
    
    # Trigger the pipeline with the updated configuration
    run = Client().trigger_pipeline(
        snapshot_name_or_id=snapshot.id,
        run_configuration=config,
    )
    
    print(f"Triggered pipeline run with ID: {run.id}")

Using the REST API:

For this you'll need a URL for a ZenML server. For those with a ZenML Pro account, you can find the URL in the dashboard in the following location:

Where to find the ZenML server URL

You can also find the URL via the CLI by running:

zenml status | grep "API:" | awk '{print $2}'

Important: Use Workspace API, Not ZenML Pro API

Snapshots are triggered via your Workspace API (your individual workspace URL), not the ZenML Pro API (cloudapi.zenml.io). Make sure you're using the correct URL from your workspace dashboard.

The REST API is ideal for external system integration, allowing you to trigger pipelines from non-Python environments:

# Step 1: Get the pipeline ID
curl -X 'GET' \
  'https://<YOUR_ZENML_SERVER>/api/v1/pipelines?name=training_pipeline' \
  -H 'accept: application/json' \
  -H 'Authorization: Bearer <YOUR_TOKEN>'

# Step 2: Get the snapshot ID using the pipeline_id
curl -X 'GET' \
  'https://<YOUR_ZENML_SERVER>/api/v1/pipeline_snapshots?pipeline=<PIPELINE_ID>' \
  -H 'accept: application/json' \
  -H 'Authorization: Bearer <YOUR_TOKEN>'

# Step 3: Trigger the pipeline with custom parameters
curl -X 'POST' \
  'https://<YOUR_ZENML_SERVER>/api/v1/pipeline_snapshots/<SNAPSHOT-ID>/runs' \
  -H 'accept: application/json' \
  -H 'Content-Type: application/json' \
  -H 'Authorization: Bearer <YOUR_TOKEN>' \
  -d '{
    "run_configuration": {
      "steps": {
        "load_data": {
          "parameters": {
            "data_url": "s3://production-bucket/latest-data.csv"
          }
        },
        "train_model": {
          "parameters": {
            "model_type": "gradient_boosting"
          }
        }
      }
    }
  }'

Note: When using the REST API, you need to specify parameters at the step level, not at the pipeline level. This matches how parameters are configured in the Python client.

Security Considerations for API Tokens

When using the REST API for external systems, proper token management is critical:

Best Practice: Use Service Accounts for Automation

For production run template triggering, always use service accounts with API keys instead of personal access tokens. Personal tokens expire after 1 hour and are tied to individual users, making them unsuitable for automation.

from zenml.client import Client

# Create a service account for automated triggers
service_account = Client().create_service_account(
    name="pipeline-trigger-service",
    description="Service account for external pipeline triggering"
)

# Generate API token with appropriate permissions
token = Client().create_service_account_token(
    service_account.id,
    name="production-trigger-token",
    description="Token for production pipeline triggers"
)

print(f"Store this token securely: {token.token}")
# Make sure to save this token value securely

Why service accounts are better for automation:

• Long-lived: Tokens don't expire automatically like user tokens (1 hour)

• Dedicated: Not tied to individual team members who might leave

• Secure: Can be granted minimal permissions needed for the task

• Traceable: Clear audit trail of which system performed actions

Use this token in your API calls, and store it securely in your external system (e.g., as a GitHub Secret, AWS Secret, or environment variable). Read more about service accounts and tokens.

Method 2: Pipeline Deployments (Open Source)

Pipeline Deployments are ZenML's built-in solution for running pipelines as long-running HTTP services. Unlike Snapshots (which launch batch jobs), Deployments create persistent web services with stable URLs that handle concurrent requests through HTTP endpoints.

Pipeline Deployments are available in both the open-source and Pro versions of ZenML. They provide a fully managed deployment experience without needing to build your own API wrapper.

Why Use Deployments?

• Persistent HTTP endpoint: A stable URL for your pipeline that stays running and handles multiple requests

• Request/response semantics: Send parameters, get pipeline outputs back as JSON

• Built-in lifecycle management: Create, update, deprovision, and delete deployments through CLI or SDK

• Multiple backends: Deploy to local, Docker, Kubernetes, GCP Cloud Run, AWS App Runner, or Hugging Face Spaces

• Customizable: Add custom endpoints, authentication, middleware, and serve static dashboards alongside your API

• Shared state: Use on_init hooks to load models once and share them across all requests

Deploying the Sample Pipeline

To deploy our sample training pipeline, first ensure it accepts parameters with default values and returns meaningful outputs. Then adapt it slightly for deployment:

from typing import Dict, Any, Union, Annotated
from zenml import pipeline, step
import numpy as np
import pandas as pd
from sklearn.ensemble import RandomForestClassifier, GradientBoostingClassifier
from sklearn.model_selection import train_test_split
from sklearn.metrics import accuracy_score

@step
def load_data(data_url: str) -> pd.DataFrame:
    np.random.seed(42)
    n_samples = 1000
    print(f"Loading data from: {data_url}")
    data = pd.DataFrame({
        'feature_1': np.random.normal(0, 1, n_samples),
        'feature_2': np.random.normal(0, 1, n_samples),
        'feature_3': np.random.normal(0, 1, n_samples),
        'target': np.random.choice([0, 1], n_samples)
    })
    return data

@step
def preprocess(data: pd.DataFrame) -> Dict[str, Any]:
    X = data.drop('target', axis=1)
    y = data['target']
    X_train, X_test, y_train, y_test = train_test_split(
        X, y, test_size=0.2, random_state=42
    )
    return {
        'X_train': X_train, 'X_test': X_test,
        'y_train': y_train, 'y_test': y_test
    }

@step
def train_and_evaluate(
    datasets: Dict[str, Any],
    model_type: str = "random_forest"
) -> Annotated[Dict[str, float], "metrics"]:
    X_train, y_train = datasets['X_train'], datasets['y_train']
    X_test, y_test = datasets['X_test'], datasets['y_test']

    if model_type == "random_forest":
        model = RandomForestClassifier(n_estimators=100, random_state=42)
    elif model_type == "gradient_boosting":
        model = GradientBoostingClassifier(random_state=42)
    else:
        raise ValueError(f"Unknown model type: {model_type}")

    model.fit(X_train, y_train)
    accuracy = accuracy_score(y_test, model.predict(X_test))
    return {'accuracy': float(accuracy), 'model_type': model_type}

@pipeline
def training_pipeline(
    data_url: str = "s3://example-bucket/data.csv",
    model_type: str = "random_forest"
) -> Dict[str, float]:
    data = load_data(data_url)
    datasets = preprocess(data)
    return train_and_evaluate(datasets, model_type)

Now deploy it:

# Using the default stack (local deployer)
zenml pipeline deploy my_module.training_pipeline --name training-service

Or via the SDK:

deployment = training_pipeline.deploy(deployment_name="training-service")
print(f"Deployment URL: {deployment.url}")

Invoking the Deployment

Once running, you can invoke the deployment from any external system:

# Via ZenML CLI
zenml deployment invoke training-service \
    --data_url="s3://production-bucket/latest-data.csv" \
    --model_type="gradient_boosting"

# Via curl (from any external system)
curl -X POST http://your-deployment-url/invoke \
  -H "Content-Type: application/json" \
  -d '{
    "parameters": {
      "data_url": "s3://production-bucket/latest-data.csv",
      "model_type": "gradient_boosting"
    }
  }'

The response includes pipeline outputs, execution metadata, and run tracking:

{
    "success": true,
    "outputs": {
        "metrics": {"accuracy": 0.875, "model_type": "gradient_boosting"}
    },
    "execution_time": 8.2,
    "metadata": {
        "deployment_name": "training-service",
        "run_id": "f2e9a3a7-afa3-459e-a970-8558358cf1fb",
        "parameters_used": {
            "data_url": "s3://production-bucket/latest-data.csv",
            "model_type": "gradient_boosting"
        }
    }
}

Production Deployment with Kubernetes

For production use, deploy to Kubernetes:

# Register a Kubernetes deployer
zenml integration install kubernetes --uv -y
zenml deployer register k8s-deployer --flavor=kubernetes \
    --kubernetes_namespace=ml-services

# Update your stack
zenml stack update -d k8s-deployer

# Deploy
zenml pipeline deploy my_module.training_pipeline --name training-service

The Kubernetes deployer provides health probes, scaling, labels/annotations, and the ability to attach additional Kubernetes resources (Ingress, HPA, etc.). See the Kubernetes deployer docs for full details.

Adding Authentication

@pipeline(
    settings={
        "deployer": {
            "generate_auth_key": True,
        }
    }
)
def training_pipeline(
    data_url: str = "s3://example-bucket/data.csv",
    model_type: str = "random_forest"
) -> Dict[str, float]:
    ...

This generates an auth key that must be included in requests:

curl -X POST http://your-deployment-url/invoke \
  -H "Authorization: Bearer <GENERATED_AUTH_KEY>" \
  -H "Content-Type: application/json" \
  -d '{"parameters": {"data_url": "s3://new-data/", "model_type": "random_forest"}}'

Managing Deployments

# List all deployments
zenml deployment list

# View deployment details and connection info
zenml deployment describe training-service

# Follow logs
zenml deployment logs training-service -f

# Update with new code
zenml pipeline deploy my_module.training_pipeline --name training-service --update

# Tear down
zenml deployment deprovision training-service

For comprehensive documentation on deployment configuration, lifecycle management, custom endpoints, middleware, and more, see the Pipeline Deployment guide.

Method 3: Building a Custom Trigger API (Open Source)

If you need full control over the triggering mechanism beyond what Pipeline Deployments offer — for example, custom routing logic, integration with a specific framework, or embedding pipeline triggers within an existing application — you can create your own API wrapper around pipeline execution.

The custom trigger API solution consists of the following components:

1. Pipeline Definition Module - Contains your pipeline code

2. FastAPI Web Server - Provides HTTP endpoints for triggering pipelines

3. Dynamic Pipeline Loading - Loads and executes pipelines on demand

4. Authentication - Secures the API with API key authentication

5. Containerization - Packages everything for deployment

Creating a Pipeline Module

First, create a module containing your pipeline definitions. This will be imported by the API service:

# common.py
from typing import Dict, Any, Union
from zenml import pipeline, step
import numpy as np
import pandas as pd
from sklearn.ensemble import RandomForestClassifier, GradientBoostingClassifier
from sklearn.model_selection import train_test_split
from sklearn.metrics import accuracy_score
from zenml.config import DockerSettings

@step
def load_data(data_url: str) -> pd.DataFrame:
    """Load data from a URL (simulated for this example)."""
    # For demonstration, we'll create synthetic data
    np.random.seed(42)
    n_samples = 1000
    
    print(f"Loading data from: {data_url}")
    # In a real scenario, you'd load from data_url
    # E.g., pd.read_csv(data_url)
    
    data = pd.DataFrame({
        "feature_1": np.random.normal(0, 1, n_samples),
        "feature_2": np.random.normal(0, 1, n_samples),
        "feature_3": np.random.normal(0, 1, n_samples),
        "target": np.random.choice([0, 1], n_samples),
    })
    return data

@step
def preprocess(data: pd.DataFrame) -> Dict[str, Any]:
    """Split data into train and test sets."""
    X = data.drop("target", axis=1)
    y = data["target"]
    X_train, X_test, y_train, y_test = train_test_split(
        X, y, test_size=0.2, random_state=42
    )
    return {
        "X_train": X_train,
        "X_test": X_test,
        "y_train": y_train,
        "y_test": y_test,
    }

@step
def train_model(
    datasets: Dict[str, Any], model_type: str = "random_forest"
) -> Union[RandomForestClassifier, GradientBoostingClassifier]:
    """Train a model based on the specified type."""
    X_train = datasets["X_train"]
    y_train = datasets["y_train"]
    
    if model_type == "random_forest":
        model = RandomForestClassifier(n_estimators=100, random_state=42)
    elif model_type == "gradient_boosting":
        model = GradientBoostingClassifier(random_state=42)
    else:
        raise ValueError(f"Unknown model type: {model_type}")
    
    print(f"Training a {model_type} model...")
    model.fit(X_train, y_train)
    return model

@step
def evaluate(
    datasets: Dict[str, Any],
    model: Union[RandomForestClassifier, GradientBoostingClassifier],
) -> Dict[str, float]:
    """Evaluate the model and return metrics."""
    X_test = datasets["X_test"]
    y_test = datasets["y_test"]
    
    y_pred = model.predict(X_test)
    accuracy = accuracy_score(y_test, y_pred)
    
    print(f"Model accuracy: {accuracy:.4f}")
    return {"accuracy": float(accuracy)}

# Define Docker settings for the pipeline
docker_settings = DockerSettings(
    requirements="requirements.txt",
    required_integrations=["sklearn"],
)

@pipeline(settings={"docker": docker_settings})
def training_pipeline(
    data_url: str = "example-data-source", model_type: str = "random_forest"
):
    """A configurable training pipeline that can be triggered externally."""
    data = load_data(data_url)
    datasets = preprocess(data)
    model = train_model(datasets, model_type)
    metrics = evaluate(datasets, model)
    return metrics

Creating a Requirements File

Create a requirements.txt file with the necessary dependencies:

# Requirements for pipeline trigger API
fastapi>=0.95.0
uvicorn>=0.21.0
requests>=2.28.0
# Core dependencies
scikit-learn>=1.0.0
pandas>=1.3.0
numpy>=1.20.0
# ZenML
zenml>=0.80.1

Creating a FastAPI Wrapper

Next, create the pipeline_api.py file with the FastAPI application:

import os
import sys
import importlib.util
from typing import Dict, Any, Optional
import threading
from fastapi import FastAPI, HTTPException, Depends, Security
from fastapi.security import APIKeyHeader
from pydantic import BaseModel
import uvicorn

# Import the training pipeline from the common module
from common import training_pipeline

# Setup FastAPI app
app = FastAPI(title="ZenML Pipeline Trigger API")

# Simple API key authentication
# This environment variable serves as a security token to protect your API endpoints
# In production, use a strong, randomly generated key stored securely
API_KEY = os.environ.get("PIPELINE_API_KEY", "your-secure-api-key")
api_key_header = APIKeyHeader(name="X-API-Key")

async def get_api_key(api_key: str = Security(api_key_header)):
    if api_key != API_KEY:
        raise HTTPException(status_code=401, detail="Invalid API key")
    return api_key

# Request model for pipeline parameters
class StepParameter(BaseModel):
    parameters: Dict[str, Any]

class PipelineRequest(BaseModel):
    pipeline_name: str
    steps: Dict[str, StepParameter] = {}
    config_path: Optional[str] = None

# Import a pipeline dynamically
def import_pipeline(pipeline_name):
    """Import a pipeline function from available modules."""
    # First try to import from known pipelines
    if pipeline_name == "training_pipeline":
        return training_pipeline
    
    # Try importing from other modules
    try:
        spec = importlib.util.find_spec("common")
        if spec is None:
            raise ImportError(f"Module 'common' not found")
            
        module = importlib.util.module_from_spec(spec)
        spec.loader.exec_module(module)
        
        if not hasattr(module, pipeline_name):
            raise AttributeError(f"Pipeline '{pipeline_name}' not found in module")
            
        return getattr(module, pipeline_name)
    except Exception as e:
        raise HTTPException(status_code=404, detail=f"Pipeline not found: {str(e)}")

@app.post("/trigger", status_code=202)
async def trigger_pipeline(
    request: PipelineRequest, 
    api_key: str = Depends(get_api_key)
):
    """Trigger a pipeline asynchronously."""
    # Start a background task and return immediately
    
    def run_pipeline():
        try:
            pipeline_func = import_pipeline(request.pipeline_name)
            if request.config_path:
                configured_pipeline = pipeline_func.with_options(
                    config_path=request.config_path
                )
            else:
                configured_pipeline = pipeline_func
                
            # Extract parameters from steps
            step_parameters = {}
            if request.steps:
                for step_name, step_config in request.steps.items():
                    if step_config.parameters:
                        step_parameters.update(step_config.parameters)
                        
            configured_pipeline(**step_parameters)
            print(f"Async pipeline '{request.pipeline_name}' completed")
        except Exception as e:
            print(f"Async pipeline '{request.pipeline_name}' failed: {str(e)}")
            
    # Start the pipeline in a background thread
    thread = threading.Thread(target=run_pipeline)
    thread.start()
    
    return {
        "status": "accepted",
        "message": "Pipeline triggered asynchronously",
    }

if __name__ == "__main__":
    print(f"Starting API server with API key: {API_KEY}")
    print("To trigger a pipeline, use:")
    print(
        'curl -X POST "http://localhost:8000/trigger" \\\n'
        '  -H "Content-Type: application/json" \\\n'
        f'  -H "X-API-Key: {API_KEY}" \\\n'
        '  -d \'{"pipeline_name": "training_pipeline", "steps": {"load_data": {"parameters": {"data_url": "custom-data-source"}}, "train_model": {"parameters": {"model_type": "gradient_boosting"}}}}\''
    )
    uvicorn.run(app, host="0.0.0.0", port=8000)

Containerizing Your API

Create a Dockerfile to containerize your API:

FROM python:3.11-slim

WORKDIR /app

# Install ZenML and other dependencies
COPY requirements.txt .
RUN pip install -U pip uv && uv pip install --system --no-cache-dir -r requirements.txt

# Copy your code
COPY . .

# Set environment variables
ENV PYTHONPATH=/app

# Define build arguments
ARG ZENML_ACTIVE_STACK_ID
ARG PIPELINE_API_KEY
ARG ZENML_STORE_URL
ARG ZENML_STORE_API_KEY

# Set environment variables from build args
ENV ZENML_ACTIVE_STACK_ID=${ZENML_ACTIVE_STACK_ID}
ENV PIPELINE_API_KEY=${PIPELINE_API_KEY}
ENV ZENML_STORE_URL=${ZENML_STORE_URL}
ENV ZENML_STORE_API_KEY=${ZENML_STORE_API_KEY}

# Export and install stack requirements
RUN if [ -n "$ZENML_ACTIVE_STACK_ID" ]; then \
    zenml stack set $ZENML_ACTIVE_STACK_ID && \
    zenml stack export-requirements $ZENML_ACTIVE_STACK_ID --output-file stack_requirements.txt && \
    uv pip install --system -r stack_requirements.txt; \
    else echo "Warning: ZENML_ACTIVE_STACK_ID not set, skipping stack requirements"; \
    fi

# Expose the port
EXPOSE 8000

# Run the API
CMD ["python", "pipeline_api.py"]

This Dockerfile includes several important features:

1. Building with the uv package installer for faster builds

2. Support for passing ZenML configuration via build arguments

3. Automatic installation of stack-specific requirements

4. Setting up environment variables for ZenML configuration

Running Your API Locally

To test the API server locally:

# Install the required dependencies
pip install -r requirements.txt

# Set the API key
export PIPELINE_API_KEY="your-secure-api-key"

# If using a remote ZenML server, set these as well
export ZENML_STORE_URL="https://your-zenml-server-url"
export ZENML_STORE_API_KEY="your-zenml-api-key"

# If you want to use a specific stack
export ZENML_ACTIVE_STACK_ID="your-stack-id"

# Start the API server
python pipeline_api.py

Deploying Your API

Build and deploy your containerized API:

# Build the Docker image
docker build -t zenml-pipeline-api \
  --build-arg ZENML_ACTIVE_STACK_ID="your-stack-id" \
  --build-arg PIPELINE_API_KEY="your-secure-api-key" \
  --build-arg ZENML_STORE_URL="https://your-zenml-server" \
  --build-arg ZENML_STORE_API_KEY="your-zenml-api-key" .

# Run the container
docker run -p 8000:8000 zenml-pipeline-api

For production deployment, you can:

• Deploy to Kubernetes with a proper Ingress and TLS

• Deploy to a cloud platform supporting Docker containers

• Set up CI/CD for automated deployments

Triggering Pipelines via the API

You can trigger pipelines through the custom API with this endpoint:

curl -X 'POST' \
  'http://your-api-server:8000/trigger' \
  -H 'accept: application/json' \
  -H 'X-API-Key: your-secure-api-key' \
  -H 'Content-Type: application/json' \
  -d '{
    "pipeline_name": "training_pipeline",
    "steps": {
      "load_data": {
        "parameters": {
          "data_url": "s3://some-bucket/new-data.csv"
        }
      },
      "train_model": {
        "parameters": {
          "model_type": "gradient_boosting"
        }
      }
    }
  }'

This method starts the pipeline in a background thread and returns immediately with a status code of 202 (Accepted), making it suitable for asynchronous execution from external systems.

Extending the API

You can extend this API to support additional features:

1. Pipeline Discovery: Add endpoints to list available pipelines

2. Run Status Tracking: Add endpoints to check the status of pipeline runs

3. Webhook Notifications: Implement callbacks when pipelines complete

4. Advanced Authentication: Implement JWT or OAuth2 for better security

5. Pipeline Scheduling: Add endpoints to schedule pipeline runs

Handling Concurrent Pipeline Execution

Important Limitation: ZenML Prevents Concurrent Pipeline Execution

ZenML's current implementation uses shared global state (like active stack and active pipeline), which prevents running multiple pipelines concurrently in the same process. If you attempt to trigger multiple pipelines simultaneously, subsequent calls will be blocked with the error:

Preventing execution of pipeline '<pipeline_name>'. If this is not intended behavior, make sure to unset the environment variable 'ZENML_PREVENT_PIPELINE_EXECUTION'.

The FastAPI example above uses threading, but due to ZenML's architecture, concurrent pipeline execution will fail. For production environments that need to handle concurrent pipeline requests, consider using Pipeline Deployments which handle this natively, or deploying your custom pipeline triggers through container orchestration platforms.

Recommended Solutions for Concurrent Execution

For production deployments, consider using:

1. Kubernetes Jobs: Deploy each pipeline execution as a separate Kubernetes Job for resource management and scaling

2. Docker Containers: Use a container orchestration platform like Docker Swarm or ECS to run separate container instances

3. Cloud Container Services: Leverage services like AWS ECS, Google Cloud Run, or Azure Container Instances

4. Serverless Functions: Deploy pipeline triggers as serverless functions (AWS Lambda, Azure Functions, etc.)

These approaches ensure each pipeline runs in its own isolated environment, avoiding the concurrency limitations of ZenML's shared state architecture.

Security Considerations

When deploying this API in production:

1. Use Strong API Keys: Generate secure, random API keys. The PIPELINE_API_KEY in the code example is a simple authentication token that protects your API endpoints. Do not use the default value in production.

2. HTTPS/TLS: Always use HTTPS for production deployments

3. Least Privilege: Use ZenML service accounts with minimal permissions

4. Rate Limiting: Implement rate limiting to prevent abuse

5. Secret Management: Use a secure secrets manager for API keys and credentials

6. Logging & Monitoring: Implement proper logging for security audits

Best Practices & Troubleshooting

Tag Snapshots

You should tag your snapshots to make them easier to find and manage. It is currently only possible using the Python SDK:

from zenml import add_tags

add_tags(tags=["my_tag"], snapshot=<SNAPSHOT-ID>)

Parameter Stability Best Practices

When triggering pipelines externally, it's crucial to maintain parameter stability to prevent unexpected behavior:

1. Document Parameter Changes: Keep a changelog of parameter modifications and their impact on pipeline behavior

2. Version Control Parameters: Store parameter configurations in version-controlled files (e.g., YAML) alongside your pipeline code

3. Validate Parameter Changes: Consider implementing validation checks to ensure new parameter values are compatible with existing pipeline steps

4. Consider Upstream Impact: Before modifying step parameters, analyze how changes might affect:

   • Downstream steps that depend on the step's output

   • Cached artifacts that might become invalid

   • Other pipelines that might be using this step

5. Use Parameter Templates: Create parameter templates for different scenarios (e.g., development, staging, production) to maintain consistency

Security Best Practices

1. API Keys: Always use API keys or tokens for authentication

2. Principle of Least Privilege: Grant only necessary permissions to service accounts

3. Key Rotation: Rotate API keys regularly

4. Secure Storage: Store credentials in secure locations (not in code)

5. TLS: Use HTTPS for all API endpoints

Monitoring and Observability

Implement monitoring for your trigger mechanisms:

import logging
from datetime import datetime

# Set up logging
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger("pipeline-trigger")

def log_trigger_attempt(pipeline_name, parameters, source):
    """Log pipeline trigger attempts."""
    timestamp = datetime.now().isoformat()
    logger.info(f"TRIGGER_ATTEMPT|{timestamp}|{pipeline_name}|{source}|{parameters}")

def log_trigger_success(pipeline_name, run_id, source):
    """Log successful pipeline triggers."""
    timestamp = datetime.now().isoformat()
    logger.info(f"TRIGGER_SUCCESS|{timestamp}|{pipeline_name}|{source}|{run_id}")

def log_trigger_failure(pipeline_name, error, source):
    """Log failed pipeline triggers."""
    timestamp = datetime.now().isoformat()
    logger.error(f"TRIGGER_FAILURE|{timestamp}|{pipeline_name}|{source}|{error}")

# Use in your trigger code
try:
    log_trigger_attempt("training_pipeline", parameters, "rest_api")
    run = Client().trigger_pipeline(
        pipeline_name_or_id="training_pipeline",
        run_configuration=run_config
    )
    log_trigger_success("training_pipeline", run.id, "rest_api")
except Exception as e:
    log_trigger_failure("training_pipeline", str(e), "rest_api")
    raise

Conclusion: Choosing the Right Approach

The best approach for triggering pipelines depends on your specific needs:

Approach	Best for	Availability
Pipeline Deployments	Persistent HTTP services for real-time inference, agents, interactive APIs. Handles concurrent requests, offers custom endpoints, auth, and full lifecycle management.	OSS and Pro
Snapshots	Batch-style pipeline runs triggered from the dashboard, SDK, CLI, or REST API. No persistent service needed.	Pro
Custom API	Full control over triggering when you need to embed pipeline triggers within an existing application or framework.	OSS and Pro

For most teams looking to expose pipelines as HTTP services, Pipeline Deployments are the recommended starting point — they provide a managed deployment experience with multiple infrastructure backends (local, Docker, Kubernetes, cloud) out of the box.

Regardless of your approach, always prioritize:

• Security (authentication and authorization)

• Reliability (error handling and retries)

• Observability (logging and monitoring)

Next Steps

Now that you understand how to trigger ZenML pipelines from external systems, consider exploring:

1. Managing scheduled pipelines for time-based execution

2. Implementing comprehensive CI/CD for your ML workflows

3. Setting up monitoring and alerting for pipeline failures