GCP PMLE · Domain 5 of 6

GCP PMLE Domain 5: Automating and orchestrating ML pipelines

Make ML repeatable: build end-to-end pipelines that validate data and models, choose an orchestrator such as Agent Platform Pipelines, Managed Service for Apache Airflow or Ray, keep preprocessing consistent, and automate retraining through CI/CD/CT.

Exam weight
~18%
Official objectives
2
Estimated questions estimate
9–11
Suggested study time estimate
12–16 hrs

About 11 min read Reviewed October 2026

Overview

Automating and orchestrating ML pipelines has two official objectives. Task 5.1 covers developing end-to-end pipelines: validating data and models, building and orchestrating with managed or unmanaged services, templates or custom code (for example Agent Platform Pipelines, Managed Service for Apache Airflow, formerly Cloud Composer, and Ray), and keeping preprocessing consistent between training and serving. Task 5.2 covers automating retraining: setting a retraining policy and deploying models through CI/CD/CT pipelines with tools such as Cloud Build.

At about 18% of the exam, this section sits in the middle by weight. It reflects MLOps practice, where the pipeline, not the notebook, is the unit that gets versioned, tested and run again.

How it's tested

Domain 5 scenarios describe a team's current manual process, its existing tools and a reliability or speed goal. You choose the orchestrator, the validation step, the trigger for retraining or the CI/CD design. The best answer is usually the most managed option that fits the team's existing stack.

  • Choose the orchestrator

    A team's tools and workload are described, and you select Agent Platform Pipelines, Managed Service for Apache Airflow, Kubeflow or Ray.

  • Add a gate

    Bad data or a weaker model reached production, and you place a validation or evaluation step that would have stopped it.

  • Retraining trigger

    Data or performance changes are described, and you choose a schedule, a monitoring-based trigger or an event trigger for retraining.

Reasoning it through

A data team already schedules dozens of ETL jobs as Airflow DAGs, and now needs a weekly step that launches model training after the ETL finishes. Building a separate Kubeflow cluster adds operations work. Cron on a VM lacks dependency handling. Managed Service for Apache Airflow can trigger an Agent Platform training pipeline as a task in the existing DAG, so extending Airflow is the expected answer.

Objectives 2

Official objectives from the Google Cloud outline, in outline order.

5.1 Developing end-to-end ML pipelinesData and model validation, managed or unmanaged orchestration, templates or custom code, and consistent preprocessing.
What it means
This task covers building pipelines that run every ML step in order. You validate data and models, build and orchestrate pipelines with managed or unmanaged services, templates or custom code, using options such as Agent Platform Pipelines, Managed Service for Apache Airflow and Ray, and make sure data preprocessing is consistent between training and serving.
What to understand
Agent Platform Pipelines runs Kubeflow Pipelines or TFX definitions serverlessly, records lineage in ML Metadata and caches unchanged steps. Managed Service for Apache Airflow (formerly Cloud Composer) suits broader workflow scheduling across data and ML jobs. Ray on Agent Platform or GKE scales Python workloads such as distributed preprocessing, tuning and training. Validation checks schema, ranges and distribution shifts in data before training, and compares a candidate model with the current one on an evaluation set before deployment. To keep preprocessing identical, package it with the model, use the BigQuery ML TRANSFORM clause, TensorFlow Transform, or shared features in Feature Store.
How it shows up
A team's stack and pipeline problem are described, and you choose the orchestrator, the validation gate or the consistency method.
Common confusion
Rewriting preprocessing separately in the serving application, which causes training-serving skew.
5.2 Automating model retrainingRetraining policy and deploying models through CI/CD/CT pipelines with tools such as Cloud Build.
What it means
This task covers keeping models fresh without manual work. You decide on a retraining policy suited to how fast the data and business change, and deploy models through continuous integration, delivery and training pipelines using services such as Cloud Build.
What to understand
Retraining can be scheduled, triggered by new data arriving, or triggered when monitoring detects drift or a drop in quality. A good policy balances cost against how quickly performance decays. Continuous integration tests pipeline code and components; continuous delivery deploys pipeline definitions and models; continuous training reruns the pipeline to produce new models. Cloud Build can run tests and submit pipelines when code changes in a repository, with artifacts stored in Artifact Registry. A new model should be promoted only after it beats the current model on agreed metrics, often with a canary rollout or human approval.
How it shows up
A data change or quality decline is described and you choose the retraining trigger or the CI/CD/CT component that should act.
Common confusion
Assuming retraining on a fixed daily schedule is always safest, even when data rarely changes.

High-yield concepts

The terms, rules and formulas this domain turns into questions most often.

  • Agent Platform Pipelines

    Serverless runner for Kubeflow Pipelines and TFX workflows.

    Remember
    Lineage, caching and managed execution.
    Tested as
    Default ML orchestrator questions.
    Trap
    Assuming you must manage a Kubeflow cluster.

  • Managed Service for Apache Airflow

    Google's managed Airflow, formerly Cloud Composer.

    Remember
    Best for cross-system workflow scheduling.
    Tested as
    Teams with existing Airflow DAGs.
    Trap
    Searching only for the old Cloud Composer name.

  • Ray

    Framework for scaling Python workloads across a cluster.

    Remember
    Available on Agent Platform and GKE.
    Tested as
    Distributed Python processing scenarios.
    Trap
    Treating it as a scheduler for ETL jobs.

  • Kubeflow Pipelines SDK

    Python SDK to define pipeline components and graphs.

    Remember
    Runs on Agent Platform Pipelines or self-managed Kubeflow.
    Tested as
    Pipeline definition questions.
    Trap
    Confusing the SDK with the runtime.

  • TFX

    TensorFlow Extended components for production pipelines.

    Remember
    Includes data validation and transform components.
    Tested as
    Validation and transform questions.
    Trap
    Assuming it works only on-premises.

  • Data validation

    Checking schema, ranges and distributions before training.

    Remember
    Stops bad data early.
    Tested as
    Pipeline gate questions.
    Trap
    Validating only after deployment.

  • Model evaluation gate

    Compare a candidate model with the current one before deploying.

    Remember
    Promote only when metrics improve.
    Tested as
    Automated deployment questions.
    Trap
    Auto-deploying every retrained model.

  • Training-serving consistency

    Same preprocessing code or features in training and serving.

    Remember
    TRANSFORM clause, TensorFlow Transform or Feature Store.
    Tested as
    Skew prevention questions.
    Trap
    Rewriting logic in the serving app.

  • Continuous training (CT)

    Automatically rerunning the training pipeline.

    Remember
    Triggered by schedule, new data or monitoring.
    Tested as
    MLOps maturity questions.
    Trap
    Confusing CT with CD.

  • Cloud Build

    Managed CI/CD service that runs build steps.

    Remember
    Tests code and submits pipelines on commit.
    Tested as
    CI/CD questions.
    Trap
    Using it as a training service.

  • Pipeline caching

    Reusing outputs of unchanged steps.

    Remember
    Saves time and cost on reruns.
    Tested as
    Efficiency questions.
    Trap
    Forgetting to disable it when inputs change outside the pipeline.

Side by side

Concepts candidates mix up, compared on the point that decides the answer.

Orchestration options

ConceptMeaningKey differenceCommon trap
Agent Platform Pipelines Managed ML pipeline execution. ML-native with lineage and caching. Using it to schedule unrelated ETL.
Managed Service for Apache Airflow Managed Airflow DAG scheduling. General workflows across many systems. Expecting built-in ML lineage.
Ray Distributed Python compute framework. Scales compute inside a step. Treating it as a full workflow scheduler.
Self-managed Kubeflow on GKE Kubeflow you install and operate. Maximum control, more operations work. Picking it when a managed option meets needs.

Retraining triggers

ConceptMeaningKey differenceCommon trap
Schedule Retrain at fixed intervals. Simple and predictable. Wasting cost when data is stable.
New data arrival Retrain when enough fresh data lands. Tied to data availability. Retraining on tiny increments.
Monitoring alert Retrain when drift or quality decline is detected. Responds to actual need. Lacking ground truth to confirm the decline.

CI, CD and CT

ConceptMeaningKey differenceCommon trap
Continuous integration Test code and components on each change. Catches broken pipeline code. Thinking it trains models.
Continuous delivery Release pipelines and models to environments. Moves validated artifacts forward. Skipping evaluation gates.
Continuous training Rerun training automatically. Produces new model versions. Deploying them without comparison.

Exam traps

Each trap is an answer that looks right. Learn the clue that should change your mind.

  1. Self-managed Kubeflow

    What candidates pick
    Install Kubeflow on a GKE cluster
    Why it looks right
    Kubeflow is the well-known ML pipeline tool.
    Why it's wrong
    Agent Platform Pipelines runs the same pipeline definitions serverlessly, avoiding cluster upkeep.

    Clue to watch for The team wants minimal infrastructure management.

  2. Auto-deploy retrained models

    What candidates pick
    Deploy every new model from the pipeline
    Why it looks right
    It keeps the model as fresh as possible.
    Why it's wrong
    A retrained model can be worse; an evaluation gate must confirm it beats the current one.

    Clue to watch for Production quality must not regress.

  3. Cron on a VM

    What candidates pick
    Schedule training with cron on a Compute Engine VM
    Why it looks right
    It is quick to set up.
    Why it's wrong
    Cron has no dependency handling, retries or lineage, which managed orchestrators provide.

    Clue to watch for Multiple dependent steps and audit needs.

  4. Daily retraining by default

    What candidates pick
    Retrain the model every day
    Why it looks right
    Fresh models feel safer.
    Why it's wrong
    If data changes slowly, daily retraining adds cost without benefit; a drift-based trigger fits better.

    Clue to watch for Data distribution is stable over months.

  5. Serving-side preprocessing rewrite

    What candidates pick
    Reimplement feature logic in the web app
    Why it looks right
    The app team owns the request path.
    Why it's wrong
    Two code paths drift apart and cause skew; ship preprocessing with the model or use shared features.

    Clue to watch for Predictions differ between offline tests and production.

  6. Ray as a scheduler

    What candidates pick
    Use Ray to schedule nightly data and training jobs
    Why it looks right
    Ray runs distributed jobs.
    Why it's wrong
    Ray scales computation but is not a workflow scheduler; Airflow or Agent Platform Pipelines handle scheduling.

    Clue to watch for The need is ordering and timing of jobs.

Answer framework

Apply these steps when a scenario asks how to automate or orchestrate ML work.

  1. Identify what already exists: an Airflow estate points to Managed Service for Apache Airflow, while a new ML workflow points to Agent Platform Pipelines.
  2. Use Ray inside a pipeline step when a Python workload needs distributed compute, not as the scheduler for the whole workflow.
  3. Add data validation before training and a model evaluation gate before deployment so bad inputs or weaker models stop early.
  4. Keep preprocessing in one place, through the TRANSFORM clause, TensorFlow Transform or Feature Store, so serving matches training.
  5. Set a retraining policy based on drift, new data or schedule, and use Cloud Build for CI/CD with canary promotion of new models.

Practice question

A team retrains a churn model weekly with Agent Platform Pipelines and deploys the output automatically. Last week a retrained model performed worse than the previous one and hurt campaign results. What change should the ML engineer make to the pipeline?

Want more like this? The GCP PMLE practice test has questions for every domain.

Checklist

0 of 11

Tick each one you can do without notes. Your ticks stay in this browser.

Study strategy

Run one real pipeline before memorizing anything. Use a Google sample to build a Kubeflow Pipelines definition with data preparation, training, evaluation and a conditional deploy step, then run it on Agent Platform Pipelines. Rerun it to see caching and open the lineage view. This single exercise covers most of task 5.1.

Next, read the overviews for Managed Service for Apache Airflow and Ray on Agent Platform, and write one sentence on when each is the better answer. Note the old Cloud Composer name so older courses still make sense. Then read Google's MLOps guidance on continuous delivery and automation pipelines, and sketch the CI, CD and CT loop with Cloud Build at the start.

For task 5.2, write a short retraining policy for three example models with different data change rates. Because this section draws on Sections 2 through 4, review it after those sections and take a timed set of mixed MLOps questions.

FAQ

How much of the GCP PMLE exam is Automating and orchestrating ML pipelines?

Automating and orchestrating ML pipelines is about 18% of the exam, roughly nine to eleven questions across two official objectives.

What is Managed Service for Apache Airflow?

It is Google's managed Apache Airflow service, previously called Cloud Composer, and the June 2026 guide uses the new name.

When should I use Agent Platform Pipelines instead of Airflow?

For ML-specific workflows that benefit from lineage, caching and managed execution, especially when no Airflow environment already exists.

What does continuous training mean?

Automatically rerunning the training pipeline on a schedule, when new data arrives or when monitoring detects a decline.

How do I stop a weaker model from deploying?

Add an evaluation step that compares the candidate with the production model and deploys only when agreed metrics improve.

Where does Ray fit?

Ray scales Python computation, such as distributed preprocessing or tuning, inside a pipeline step on Agent Platform or GKE.

Other Artificial Intelligence & Machine Learning certifications

Where the same subject shows up in other exams. Cards marked "Same subject" open the matching domain.

Test your Domain 5 knowledge

You have reviewed pipeline orchestration, validation gates and retraining automation. Check whether you can choose the right orchestrator and trigger in timed scenarios.

Take the GCP PMLE practice test Check your GCP PMLE readiness

Keep going