Vertex AI Pipelines / MLOps Complete Guide: Kubeflow, CT, and Model Monitoring

Vertex AI Pipelines is GCP's MLOps backbone, providing a managed runtime for Kubeflow Pipelines (KFP) v2 and TFX. Combined with Continuous Training (CT), Model Registry, Model Monitoring, and Feature Store, it forms the foundation of a complete MLOps stack.

MLOps Maturity Levels

KFP v2 Pipeline Example

from kfp import dsl, compiler
from kfp.dsl import component, pipeline, Input, Output, Model, Dataset

@component(base_image="python:3.11", packages_to_install=["pandas", "scikit-learn"])
def load_data(output_data: Output[Dataset]):
    import pandas as pd
    df = pd.read_csv("gs://my-bucket/train.csv")
    df.to_csv(output_data.path, index=False)

@component(packages_to_install=["scikit-learn", "joblib"])
def train(data: Input[Dataset], model: Output[Model]):
    import pandas as pd, joblib
    from sklearn.linear_model import LogisticRegression
    df = pd.read_csv(data.path)
    clf = LogisticRegression().fit(df.drop("y", axis=1), df["y"])
    joblib.dump(clf, model.path)

@pipeline(name="train-pipeline")
def my_pipeline():
    data_op = load_data()
    train_op = train(data=data_op.outputs["output_data"])

compiler.Compiler().compile(my_pipeline, "pipeline.json")

# Run
from google.cloud import aiplatform
aiplatform.PipelineJob(
    display_name="my-run",
    template_path="pipeline.json",
    pipeline_root="gs://my-bucket/pipelines",
).run()

Continuous Training (CT) Pattern

1. Cloud Scheduler publishes a weekly message to Pub/Sub
2. Cloud Functions kicks off a Vertex AI Pipelines run
3. Latest features are pulled from Feature Store
4. Training and evaluation execute
5. If metrics clear the threshold, the model is registered in Model Registry
6. Model Monitoring continuously watches for drift
7. Detected drift also triggers a new Pipelines run

Model Registry

• Model versioning with aliases (champion / challenger)
• Model cards (metadata and intended use)
• Endpoint deployment (one-click promotion to production)
• Lineage tracking across BigQuery, Cloud Storage, and Feature Store

Model Monitoring

• Feature Skew: distribution gap between training and production data
• Feature Drift: drift in production data over time
• Prediction Drift: shift in the distribution of predicted values
• Data Quality: NULL and anomaly detection
• Pub/Sub notifications wired into a CT retraining pipeline

Pricing Examples

Comparison with Other MLOps Platforms

Best Practices

• Design components to be reusable and generic
• Centralize feature management in Feature Store
• Swap production models via Model Registry aliases
• Trigger pipeline updates automatically through Cloud Build
• Track experiments with Vertex AI Experiments
• Visualize runs with Vertex AI TensorBoard
• Pin Pipeline Root to GCS and apply lifecycle rules

* Google Cloud and Vertex AI are trademarks of Google LLC. For the latest information, see the official Vertex AI Pipelines docs.