| Workflow Orchestration | Built around defining and running end-to-end ML/AI pipelines with steps, artifacts, and repeatable executions across environments | Orchestrates serving resources on Kubernetes, not training or evaluation workflows — no native pipeline or DAG system |
| Integration Flexibility | Composable stack with 50+ MLOps integrations — swap orchestrators, trackers, and deployers without code changes | Deep integration with K8s serving runtimes, but scoped to inference — doesn't integrate across the broader ML toolchain |
| Vendor Lock-In | Cloud-agnostic by design — stacks make it easy to switch infrastructure providers and tools as needs change | Open-source CNCF project not tied to any cloud — lock-in is to Kubernetes itself and optionally Knative/Istio for key features |
| Setup Complexity | pip install zenml — start locally and progressively adopt infrastructure via stack components without needing K8s on day one | Requires installing Kubernetes controllers, CRDs, and optionally Knative and networking dependencies for full feature set |
| Learning Curve | Python-first abstraction matches how ML engineers write training code, with infrastructure details pushed into configuration | Approachable for K8s-native teams but demands comfort with CRDs, cluster networking, and serving runtime concepts |
| Scalability | Delegates compute to scalable backends — Kubernetes, Spark, cloud ML services — for unlimited horizontal scaling | Designed for scalable multi-tenant inference with request-based autoscaling, scale-to-zero, and canary rollout patterns |
| Cost Model | Open-source core is free — pay only for your own infrastructure, with optional managed cloud for enterprise features | Apache-2.0 with no per-seat or per-request fees — costs are infrastructure and operations, with scale-to-zero reducing waste |
| Collaboration | Code-native collaboration through Git, CI/CD, and code review — ZenML Pro adds RBAC, workspaces, and team dashboards | Relies on Kubernetes-native collaboration (GitOps, cluster RBAC) — no ML-specific collaboration layer for experiments or artifacts |
| ML Frameworks | Use any Python ML framework — TensorFlow, PyTorch, scikit-learn, XGBoost, LightGBM — with native materializers and tracking | Multi-framework serving support (TensorFlow, PyTorch, scikit-learn, XGBoost, ONNX) plus growing GenAI/LLM runtimes |
| Monitoring | Integrates Evidently, WhyLogs, and other monitoring tools as stack components for automated drift detection and alerting | Serving-time metrics via Prometheus, payload logging, and drift/outlier detection integrations — scoped to inference endpoints |
| Governance | ZenML Pro provides RBAC, SSO, workspaces, and audit trails — self-hosted option keeps all data in your own infrastructure | Governance inherited from Kubernetes (RBAC, namespaces) — ML-specific governance like training-to-deploy lineage is out of scope |
| Experiment Tracking | Native metadata tracking plus seamless integration with MLflow, Weights & Biases, Neptune, and Comet for rich experiment comparison | Does not track experiments — serves whatever model artifact is provided and exposes runtime-level status and metrics |
| Reproducibility | Automatic artifact versioning, code-to-Git linking, and containerized execution guarantee reproducible pipeline runs | Serving configs are reproducible via K8s manifests, but end-to-end reproducibility of training data, code, and environments is out of scope |
| Auto-Retraining | Schedule pipelines via any orchestrator or use ZenML Pro event triggers for drift-based automated retraining workflows | Can roll out new model versions once produced, but does not automate retraining or the upstream triggers that decide when to retrain |
from zenml import pipeline, step, Model
from zenml.integrations.mlflow.steps import (
mlflow_model_deployer_step,
)
import pandas as pd
from sklearn.ensemble import RandomForestRegressor
from sklearn.metrics import mean_squared_error
import numpy as np
@step
def ingest_data() -> pd.DataFrame:
return pd.read_csv("data/dataset.csv")
@step
def train_model(df: pd.DataFrame) -> RandomForestRegressor:
X, y = df.drop("target", axis=1), df["target"]
model = RandomForestRegressor(n_estimators=100)
model.fit(X, y)
return model
@step
def evaluate(model: RandomForestRegressor, df: pd.DataFrame) -> float:
X, y = df.drop("target", axis=1), df["target"]
preds = model.predict(X)
return float(np.sqrt(mean_squared_error(y, preds)))
@step
def check_drift(df: pd.DataFrame) -> bool:
# Plug in Evidently, Great Expectations, etc.
return detect_drift(df)
@pipeline(model=Model(name="my_model"))
def ml_pipeline():
df = ingest_data()
model = train_model(df)
rmse = evaluate(model, df)
drift = check_drift(df)
# Runs on any orchestrator, logs to MLflow,
# tracks artifacts, and triggers retraining — all
# in one portable, version-controlled pipeline
ml_pipeline()import asyncio
from kubernetes import client
from kserve import (
KServeClient, constants,
V1beta1InferenceService, V1beta1InferenceServiceSpec,
V1beta1PredictorSpec, V1beta1SKLearnSpec,
RESTConfig, InferenceRESTClient,
)
async def main():
name, namespace = "sklearn-iris", "kserve-test"
isvc = V1beta1InferenceService(
api_version=constants.KSERVE_V1BETA1,
kind=constants.KSERVE_KIND,
metadata=client.V1ObjectMeta(
name=name, namespace=namespace
),
spec=V1beta1InferenceServiceSpec(
predictor=V1beta1PredictorSpec(
sklearn=V1beta1SKLearnSpec(
storage_uri="s3://my-bucket/iris/model.joblib"
)
)
),
)
ksc = KServeClient()
ksc.create(isvc)
ksc.wait_isvc_ready(name, namespace=namespace)
rest = InferenceRESTClient(
RESTConfig(protocol="v1", retries=5, timeout=30)
)
result = await rest.infer(
f"http://{name}.{namespace}",
{"instances": [[6.8, 2.8, 4.8, 1.4]]},
model_name=name,
)
print(result)
asyncio.run(main())
# Deploys and serves a model on Kubernetes.
# No pipeline orchestration, experiment tracking,
# artifact lineage, or retraining automation.
# Focused purely on the inference endpoint.
.png)
