Zenml Pipeline Authoring

Author ZenML Pipelines

This skill guides pipeline authoring: steps, artifacts, configuration, Docker settings, materializers, metadata, secrets, and visualizations.

Start Here: Interview the User

Do not rush to code. Before writing a single line, thoroughly understand what the user wants to build. The interview is the most important step — a well-scoped pipeline that does 3 things well beats a sprawling one that does 10 things poorly.

For complex or multi-pipeline projects: If the user describes something ambitious (e.g., "build me an end-to-end ML platform with data ingestion, feature engineering, training, evaluation, deployment, monitoring, and retraining"), or if they mention multiple pipelines, invoke the zenml-scoping skill first. It runs a deeper architectural interview that decomposes the system into pipeline units, identifies what doesn't belong in a pipeline at all, and produces a pipeline_architecture.md spec. Once that's done, come back here to build each pipeline one at a time.

For single, focused pipelines: If the user's request is clearly one pipeline (e.g., "build a training pipeline for my CSV data"), proceed with the questions below. If the answers are obvious from context, infer them and proceed. Only ask when genuinely ambiguous.

Q1: Static or dynamic pipeline? Most pipelines are static (fixed DAG). Use dynamic (@pipeline(dynamic=True)) only when the number of steps or their wiring depends on runtime values (e.g., "process N documents where N comes from a query"). See Dynamic Pipelines and references/dynamic-pipelines.md.

Q2: Local or remote orchestrator? If remote (Kubernetes, Vertex AI, SageMaker, AzureML), the Artifact Golden Rule is critical, and you will need Docker Settings. If local-only for now, you can defer those concerns. Ask whether the user already has a stack set up — if not, point them to the ZenML docs for stack setup (this skill does not cover stack creation).

Q3: Any custom Python types? If steps produce or consume types beyond builtins, pandas, numpy, or Pydantic models, you likely need a custom materializer. Note: Pydantic BaseModel subclasses have a built-in materializer — often the simplest alternative to writing a custom materializer.

Q4: Where should the project live? Ask the user where to create the project — a new subfolder, or the current directory. If the current directory is not empty, suggest a new subfolder.

Q5: What are the data sources? Understand where data comes from: local CSV/Parquet files, a database (Snowflake, PostgreSQL), an API, cloud storage? This determines the first step's implementation and whether secrets are needed. If credentials are involved, always use ZenML Secrets — never pass passwords as CLI arguments or in config files.

Q6: Does the user want a small-data development mode? Many users want to iterate quickly with a fraction of the dataset. Plan for a --sample-size or --small CLI flag in run.py.

Core Anatomy

Defining steps

A step is a Python function decorated with @step. Type hints on inputs and outputs are required — they control serialization, caching, and dashboard display.

from zenml import step

@step
def train_model(X_train: pd.DataFrame, lr: float = 0.01) -> sklearn.base.BaseEstimator:
    """lr is a parameter (literal value); X_train is an artifact (from upstream step)."""
    model = LogisticRegression(C=1/lr).fit(X_train.drop("target", axis=1), X_train["target"])
    return model

Parameters vs artifacts: If a step input comes from another step's output, it is an artifact. If it is a literal value passed directly (JSON-serializable), it is a parameter. ZenML handles them differently.

Named and multi-output steps

Use Annotated to give outputs stable names. Use Tuple for multiple outputs:

from typing import Annotated, Tuple
from zenml import step
import pandas as pd

@step
def split_data(df: pd.DataFrame, ratio: float = 0.8) -> Tuple[
    Annotated[pd.DataFrame, "train"],
    Annotated[pd.DataFrame, "test"],
]:
    idx = int(len(df) * ratio)
    return df.iloc[:idx], df.iloc[idx:]

Wiring a pipeline

from zenml import pipeline

@pipeline
def training_pipeline(dataset_path: str = "data.csv", lr: float = 0.01) -> None:
    df = load_data(path=dataset_path)
    train, test = split_data(df=df)
    model = train_model(X_train=train, lr=lr)
    evaluate(model=model, X_test=test)

if __name__ == "__main__":
    training_pipeline()

Pipeline parameters (like dataset_path) can be overridden at runtime or via YAML config.

Step invocation IDs

When you call a step multiple times in one pipeline, ZenML auto-suffixes the name (scale, scale_2). Override with my_step(id="custom_id").

Project structure

Every pipeline project MUST follow this layout. This is non-negotiable — it produces clean, maintainable projects:

my_pipeline_project/
├── steps/                    # One file per step
│   ├── load_data.py
│   ├── preprocess.py
│   ├── train_model.py
│   └── evaluate.py
├── pipelines/
│   └── training.py           # Pipeline definition(s)
├── materializers/            # Custom materializers (if any)
│   └── my_data_materializer.py
├── visualizations/           # HTML/CSS templates for dashboard visualizations
│   └── metrics_report.html
├── configs/                  # One YAML config per environment
│   ├── dev.yaml
│   ├── staging.yaml
│   └── prod.yaml
├── run.py                    # CLI entry point (argparse, not click)
├── README.md                 # How to run, what stacks to use, etc.
└── pyproject.toml            # Dependencies — always pyproject.toml, not requirements.txt

Key rules:

• One step per file in a steps/ directory — not all steps in one steps.py.
• Separate pipeline definition from execution — pipeline in pipelines/, execution in run.py.
• Always create a README.md (not summary.md) explaining how to run the pipeline, what stacks it supports, and any setup needed. Link to the relevant ZenML docs pages (e.g., dynamic pipelines docs) rather than embedding lengthy explanations. Do NOT include stack registration or setup instructions — just say "assumes you have a ZenML stack configured" and link to https://docs.zenml.io for stack setup.
• Always use pyproject.toml for dependency declarations. Do NOT create requirements.txt alongside it — use one or the other, and pyproject.toml is the right choice.
• run.py uses argparse (not click) — click can conflict with ZenML's own click dependency.
• Run zenml init at the project root to set the source root explicitly — this prevents import failures when code runs inside containers.
• For uv/pyproject.toml package discovery, package data, and remote wheel inclusion pitfalls, see references/runtime-portability-and-approvals.md.

pyproject.toml template

Always use this as the starting point for pyproject.toml:

[project]
name = "my-pipeline-project"
version = "0.1.0"
requires-python = ">=3.12"
dependencies = [
    "zenml>=0.93",
    "pandas>=2.0",
    # Add pipeline-specific dependencies here
]

[project.optional-dependencies]
dev = [
    "pytest",
    "ruff",
    "mypy",
]

Key version constraints:

• Python >= 3.12 — ZenML's modern features and type annotations benefit from 3.12+.
• ZenML >= 0.93 — this is the minimum for current features. For dynamic pipelines, require >= 0.91 at absolute minimum (but 0.93 is safer).
• Don't pin dev tool versions (pytest, ruff, mypy) — just list them without version constraints so users get the latest.
• Prefer uv in README instructions: uv pip install -e ".[dev]" for faster and more reliable resolution. If uv is unavailable in the user's environment, use pip install -e ".[dev]".

README template notes

The README should include:

• How to install: Prefer uv pip install -e ".[dev]" and zenml integration install <name> --uv when supported. If uv is unavailable, use pip equivalents. Omit -y so users can review prompts.
• How to run: python run.py --config configs/dev.yaml
• What stacks it supports (just name them, don't explain how to register them)
• Link to specific orchestrator docs — not just generic https://docs.zenml.io. For example, if targeting Vertex AI, link to the Vertex AI orchestrator page and the GCP service connector page. Encourage users to use service connectors for authentication rather than manual credential management.
• A simple ASCII DAG visualization of the pipeline flow is a nice touch:

  load_data --> preprocess --> train_model --> evaluate
  

run.py CLI template

Every run.py should offer these flags:

import argparse
from pipelines.training import training_pipeline

def main():
    parser = argparse.ArgumentParser(description="Run the training pipeline")
    parser.add_argument("--config", default="configs/dev.yaml", help="Path to YAML config")
    parser.add_argument("--no-cache", action="store_true", help="Disable caching")
    parser.add_argument("--sample-size", type=int, default=None,
                        help="Use only N rows (for quick local iteration)")
    args = parser.parse_args()

    pipeline_instance = training_pipeline.with_options(
        config_path=args.config,
        enable_cache=not args.no_cache,
    )
    pipeline_instance(sample_size=args.sample_size)

if __name__ == "__main__":
    main()

The sample_size parameter is passed as a pipeline parameter so the data-loading step can slice the dataset.

The Artifact Golden Rule

Data must enter and move through the pipeline as artifacts, not as local file paths.

This is the single most important concept for cloud portability. When running on a remote orchestrator, each step runs in a separate container on a separate machine. There is no shared filesystem between steps.

What goes wrong

# ANTI-PATTERN: works locally, fails on cloud
@step
def preprocess(input_path: str) -> str:
    df = pd.read_csv(input_path)           # Reads from local disk
    output_path = "/tmp/processed.csv"
    df.to_csv(output_path)
    return output_path                      # Next step can't access /tmp on a different pod

@step
def train(data_path: str) -> None:
    df = pd.read_csv(data_path)             # FileNotFoundError on cloud!

The correct pattern

# CORRECT: data flows as artifacts
@step
def preprocess(input_path: str) -> pd.DataFrame:
    return pd.read_csv(input_path)          # ZenML serializes the DataFrame to the artifact store

@step
def train(data: pd.DataFrame) -> None:
    ...                                     # ZenML loads it from the artifact store — works everywhere

The first step in a pipeline is typically the one that bridges external data into the artifact world. All downstream steps receive artifacts, never file paths.

Dynamic Pipelines

Use dynamic pipelines when the DAG shape depends on runtime values. They are experimental and have restricted orchestrator support (Local, LocalDocker, Kubernetes, Vertex, SageMaker, AzureML). Always link to the dynamic pipelines documentation (https://docs.zenml.io/how-to/steps-pipelines/dynamic-pipelines) in the README since these APIs can be tricky to get right.

Minimal example

from zenml import pipeline, step

@step
def get_count() -> int:
    return 3

@step
def process(index: int) -> None:
    print(f"Processing {index}")

@pipeline(dynamic=True)
def my_dynamic_pipeline() -> None:
    count = get_count()
    count_data = count.load()       # .load() gets actual Python value
    for idx in range(count_data):
        process(index=idx)

The critical distinction: .load() vs .chunk()

You typically need both: .load() to iterate/decide, .chunk() to wire the DAG:

items = produce_list()
for i, val in enumerate(items.load()):   # load to iterate
    if val > threshold:
        chunk = items.chunk(index=i)      # chunk to wire
        process(chunk)

Fan-out with .map() and parallel execution with .submit()

For map/reduce patterns, .map() fans out over a collection. For explicit parallelism, .submit() returns a future.

See references/dynamic-pipelines.md for the complete API: .map(), .product(), .submit(), unmapped(), .unpack(), child pipelines, .embed(), runtime modes, execution-mode caveats, orchestrator support table, and limitations.

Injecting External Data

When data originates outside the pipeline (a local file, a database, an API), you need to bridge it into the artifact system.

Pattern A: ExternalArtifact(value=...)

Upload data inline when defining the pipeline. Simple but disables caching for the consuming step:

from zenml import ExternalArtifact, pipeline, step
import pandas as pd

@step
def train(data: pd.DataFrame) -> None:
    ...

@pipeline
def my_pipeline() -> None:
    df = pd.read_csv("local_data.csv")
    train(data=ExternalArtifact(value=df))

Pattern B: Pre-upload + UUID reference (for remote orchestrators)

For dynamic pipelines on remote orchestrators, the pipeline function runs inside the orchestrator pod — it cannot read your local filesystem. Pre-upload the data, then reference it by UUID:

# run.py (client-side, runs on your machine)
from zenml.artifacts.utils import save_artifact
import pandas as pd

df = pd.read_csv("local_data.csv")
art = save_artifact(data=df, name="my_dataset")
print(art.id)  # Pass this UUID to the pipeline

# pipeline.py (runs inside the orchestrator pod)
from uuid import UUID
from zenml.client import Client

@pipeline
def my_pipeline(dataset_id: str) -> None:
    artifact = Client().get_artifact_version(UUID(dataset_id))
    train(data=artifact)

Important: do not construct ExternalArtifact(id=...) in user code. The public ExternalArtifact class is for value=... uploads; its internal config carries an ID only after upload. For existing artifacts, use Client().get_artifact_version(...). The stale patterns are ExternalArtifact(name=...), version=..., and model=....

See references/external-data.md for additional patterns including register_artifact().

YAML Configuration

Separate environment-specific settings from pipeline code using YAML config files.

Minimal example

# configs/dev.yaml
enable_cache: false
parameters:
  dataset_path: "data/small.csv"
  lr: 0.05
steps:
  train_model:
    settings:
      resources:
        cpu_count: 2

training_pipeline.with_options(config_path="configs/dev.yaml")()

Configuration precedence (highest to lowest): Runtime Python code > Step-level YAML > Pipeline-level YAML > Defaults.

Always use separate config files per environment (configs/dev.yaml, configs/staging.yaml, configs/prod.yaml) — never a single config.yaml. Generate a template with zenml pipeline build-configuration my_pipeline > config_template.yaml.

Prefer with_options() (returns a copy) over configure() (mutates in place).

See references/yaml-config.md for the complete YAML schema and multi-env pattern.

Docker Settings

When running on remote orchestrators, ZenML builds Docker images for each step. Use DockerSettings to control what goes into those images.

Common patterns

from zenml.config import DockerSettings

# Install pip packages
docker = DockerSettings(
    requirements=["scikit-learn>=1.0", "pandas>=2.0"],
    apt_packages=["libgomp1"],
    environment={"PYTHONUNBUFFERED": "1"},
)

@pipeline(settings={"docker": docker})
def my_pipeline() -> None:
    ...

Step-level overrides

Apply different settings per step with @step(settings={"docker": DockerSettings(parent_image="pytorch/pytorch:2.2.0-cuda11.8-cudnn8-runtime", requirements=["transformers"])}).

Key fields: requirements (pip packages), required_integrations (ZenML integrations), apt_packages (system packages), environment (build-time env vars), runtime_environment (runtime env vars), parent_image (custom base image), python_package_installer ("uv" default or "pip"), prevent_build_reuse (force fresh builds for debugging).

See references/docker-settings.md for the full field catalog and YAML equivalents. For Apple Silicon/ARM development targeting AMD64 Linux/Kubernetes, remote package inclusion, and Kubernetes pod troubleshooting, see references/runtime-portability-and-approvals.md and the Kubernetes troubleshooting section.

Resource Settings and Resource Pools

Use ResourceSettings when a step needs explicit compute: CPUs, memory, GPUs, preemptibility, or custom pool resources.

from zenml import step
from zenml.config import ResourceSettings

@step(settings={"resources": ResourceSettings(cpu_count=4, gpu_count=1, memory="16GiB")})
def train_model(...) -> ...:
    ...

For ZenML Pro resource pools, these same step-level settings become server-side resource requests for dynamic pipelines: gpu_count maps to gpu, cpu_count maps to mcpu, memory maps to memory_mb, and ZenML adds an implicit step_run slot. Resource pools are a paid Pro feature and do not apply to static pipelines today.

Do not assume generic ResourceSettings controls every backend's machine shape. Some orchestrators or step operators expose dedicated settings for their compute layer.

Metadata Logging

Log metadata from within steps to track metrics, parameters, and results in the ZenML dashboard. This is essential for production pipelines — every training step, evaluation step, and data quality step should log relevant metadata.

Basic usage (inside a step)

from zenml import step
from zenml.utils.metadata_utils import log_metadata

@step
def train_model(X_train: pd.DataFrame, y_train: pd.Series) -> sklearn.base.BaseEstimator:
    model = LogisticRegression().fit(X_train, y_train)
    log_metadata({"accuracy": model.score(X_train, y_train), "n_samples": len(X_train)})
    return model

When called with only metadata inside a step, it automatically attaches to the current step.

Nested metadata (creates separate cards in the dashboard)

Use nested dicts — each top-level key becomes its own card: log_metadata({"model_metrics": {"accuracy": 0.95, "f1": 0.90}, "data_stats": {"n_samples": 5000}}).

Log metadata in every data loading step (row count, source), preprocessing step (rows removed), training step (all metrics, hyperparams), and evaluation step (test metrics, thresholds).

Secrets Management

When a pipeline needs credentials (database passwords, API keys, cloud tokens), never pass them as CLI arguments, config file values, or environment variables in code. Use the ZenML secret store:

# One-time setup (CLI)
zenml secret create db_credentials --host=db.example.com --username=admin --password=secret123

from zenml import step
from zenml.client import Client

@step
def load_from_database(query: str) -> pd.DataFrame:
    secret = Client().get_secret("db_credentials")
    host = secret.secret_values["host"]
    username = secret.secret_values["username"]
    password = secret.secret_values["password"]
    # Use credentials to connect...

This works on any orchestrator — the secret store is centralized in the ZenML server.

If private and public secrets can share the same name, be explicit instead of relying on search order:

private_secret = Client().get_secret("db_credentials", private=True)
shared_secret = Client().get_secret("shared_service", private=False)

Private secrets are only visible to their creator; public secrets follow the workspace/RBAC rules.

Custom Types and Materializers

When do you need a custom materializer?

Zero-effort visualizations

Return special string types for instant dashboard visualizations — no materializer needed:

from zenml.types import HTMLString, MarkdownString, CSVString

@step
def report(metrics: dict) -> HTMLString:
    # For polished HTML, load from a separate template file rather than inline strings
    html_path = Path(__file__).parent.parent / "visualizations" / "metrics_report.html"
    template = html_path.read_text()
    return HTMLString(template.format(**metrics))

Visualization quality: Keep HTML templates in a visualizations/ directory as separate .html files (optionally with embedded CSS). This makes them easy to edit and preview in a browser. Avoid writing raw HTML strings inside Python — it produces ugly, hard-to-maintain visualizations. Include proper CSS styling for a polished dashboard appearance.

Minimal custom materializer

import json, os
from typing import Any, Type
from zenml.enums import ArtifactType
from zenml.materializers.base_materializer import BaseMaterializer

class MyDataMaterializer(BaseMaterializer):
    ASSOCIATED_TYPES = (MyData,)
    ASSOCIATED_ARTIFACT_TYPE = ArtifactType.DATA

    def save(self, data: MyData) -> None:
        path = os.path.join(self.uri, "data.json")
        with self.artifact_store.open(path, "w") as f:   # Always use self.artifact_store.open()
            json.dump(data.to_dict(), f)

    def load(self, data_type: Type[Any]) -> MyData:
        path = os.path.join(self.uri, "data.json")
        with self.artifact_store.open(path, "r") as f:
            return MyData(**json.load(f))

Use self.artifact_store.open() (not plain open()) so it works with S3, GCS, and Azure Blob — not just local filesystems.

Assign it: @step(output_materializers=MyDataMaterializer).

See references/materializers.md for visualizations, metadata extraction, and registration options.

Logging

ZenML automatically captures print() and standard logging module output from within steps — no special setup needed. Logs are stored in the artifact store and visible in the dashboard.

import logging
from zenml import step

logger = logging.getLogger(__name__)

@step
def train_model(data: pd.DataFrame) -> Model:
    logger.info(f"Training on {len(data)} samples")  # Captured by ZenML
    print("Starting training...")                      # Also captured
    ...

Key points:

• Only logs from the main thread are captured. If you spawn threads, use contextvars.copy_context() to propagate ZenML's logging context.
• For remote runs, set logging verbosity via DockerSettings: DockerSettings(environment={"ZENML_LOGGING_VERBOSITY": "DEBUG"}).
• Disable step log storage with @step(enable_step_logs=False) or ZENML_DISABLE_STEP_LOGS_STORAGE=true.
• To view logs in the dashboard with a remote artifact store, you need a configured service connector.

Retry Configuration

For flaky external services or transient cloud errors, use StepRetryConfig(max_retries=3, delay=10, backoff=2) on @step(retry=...) or @pipeline(retry=...). YAML equivalent uses retry: at step or top level. To run code on failure or success, see Lifecycle Hooks.

Post-Creation Enhancements

After the pipeline is scaffolded and working, offer these enhancements. Each is optional — ask the user which they want before adding them.

Tags (add by default)

Always add tags to pipelines and key artifacts — they make filtering and organization much easier in the dashboard. Use @pipeline(tags=["training", "v1"]) and ArtifactConfig(tags=["dataset"]) on step outputs. Cascade tags (Tag(name="experiment-42", cascade=True)) automatically propagate to all artifact versions created during the run.

Model Control Plane

Track the pipeline's artifacts under a named model for versioning, promotion, and cross-pipeline artifact sharing. Use @pipeline(model=Model(name="my_model", tags=["classification"])). This enables model promotion (staging → production) and artifact exchange between training and inference pipelines. See references/post-creation.md for full patterns. For human approval gates before promotion, see references/runtime-portability-and-approvals.md.

Scheduling

Run the pipeline on a recurring schedule. Use Schedule(cron_expression="0 2 * * *") or Schedule(interval_second=3600) with pipeline.with_options(schedule=schedule). Not all orchestrators support scheduling — Kubernetes, Vertex AI, SageMaker, AzureML, Airflow, and Kubeflow do; Local and SkyPilot do not. See references/post-creation.md for the orchestrator support table and management commands.

Pipeline Deployment (HTTP serving)

For real-time inference or agent workflows, pipelines can be deployed as persistent HTTP services using pipeline.deploy(deployment_name="my_service"). This replaces the deprecated model deployer components. Use pipeline-level DeploymentSettings for ASGI app/server behavior such as custom endpoints, CORS, middleware, secure headers, thread pools, and uvicorn configuration.

Live Streaming Events

When a running step needs to surface progress before it returns — for example LLM tokens, progress updates, or live dashboard events — use zenml.streaming.publish(). Treat streaming events as best-effort live telemetry, not storage: payloads are JSON dicts, size-capped, can be dropped under load, and require server-side streaming support. Use metadata or artifacts for durable records.

Lifecycle Hooks

Run custom code at lifecycle points of a step or run with on_start, on_success, on_failure, on_end (plus on_pause/on_resume at run scope on dynamic pipelines). The common case is notifying on failure: @step(on_failure=alerter_failure_hook) posts to the active stack's alerter. on_failure and on_end optionally receive the BaseException. The others take no arguments.

One trap to respect: @pipeline(on_*=...) behaves differently on static vs dynamic pipelines. On a static pipeline it is a per-step default that each step inherits, and it never fires once at the run level. On a dynamic pipeline it fires once at the run level. To notify once per run on a static pipeline, attach the hook to a single terminal step.

run_hook(func, ...) records an arbitrary CUSTOM invocation from inside a step or dynamic pipeline, useful for instrumenting agent tool/model calls. Each lifecycle and custom firing is recorded as a queryable HookInvocation.

See references/hooks.md for the full lifecycle table, signatures, run_hook, on_init/on_cleanup, and invocation querying.

See references/post-creation.md for detailed patterns for tags, Model Control Plane, scheduling, and deployment.

Common Anti-Patterns

Resources

Reference files (detailed guides)

• references/dynamic-pipelines.md — Complete dynamic pipeline API
• references/external-data.md — Data injection patterns
• references/docker-settings.md — Full DockerSettings field catalog
• references/materializers.md — Materializer authoring guide
• references/yaml-config.md — Complete YAML config schema
• references/hooks.md — Lifecycle hooks, signatures, run_hook, invocation querying
• references/post-creation.md — Tags, Model Control Plane, scheduling, deployment
• references/runtime-portability-and-approvals.md — uv/package discovery, Apple Silicon → AMD64 builds, Kubernetes troubleshooting, approval gates

ZenML documentation

For topics not covered here (stack setup, experiment tracking, advanced deployment configuration), query the ZenML docs at https://docs.zenml.io.

When linking to docs in generated READMEs, link to specific pages rather than the generic homepage. Common links to include:

• Dynamic pipelines: https://docs.zenml.io/how-to/steps-pipelines/dynamic-pipelines
• Service connectors (for cloud auth): https://docs.zenml.io/how-to/infrastructure-deployment/auth-management
• Orchestrator-specific pages (e.g., Vertex AI, Kubernetes, SageMaker) — search docs for the specific orchestrator name
• Encourage service connectors over manual credential management for cloud stacks