PMLE Automating and Orchestrating ML Pipelines Practice Test 3 — 15 Questions

Question 1

A data science team uses Cloud Composer to orchestrate ML workflows. They need to trigger a Vertex AI pipeline after a BigQuery data load completes, and then run a Dataflow job. Which Airflow operator should they use to launch the Vertex AI pipeline?

Choose an answer to begin — your selection is scored in the full session.

Accepted Answer

VertexAIPipelineJobOperator. The VertexAIPipelineJobOperator is specifically designed to trigger a Vertex AI pipeline run from within an Airflow DAG. This operator directly corresponds to the requirement of launching a Vertex AI pipeline after a BigQuery data load completes, making it the appropriate choice for orchestrating ML workflows with Cloud Composer.

Answer

DataflowStartFlexTemplateJobOperator

Answer

VertexAICreateCustomJobOperator

Answer

MLEngineTrainingOperator

Question 2

A machine learning pipeline includes a conditional branch: if model accuracy exceeds 0.95, deploy to production; otherwise, send a notification. Which KFP SDK feature allows implementing this logic within the pipeline definition?

Accepted Answer

dsl.If. C is correct because KFP SDK's `dsl.If` is the native construct for implementing conditional branching within a pipeline definition, allowing you to conditionally execute a task (like deploying to production) based on the output of a previous component (e.g., model accuracy > 0.95). It works by creating a `Condition` object that wraps the downstream tasks, ensuring the pipeline DAG is correctly compiled and executed only when the condition evaluates to true.

Answer

dsl.ParallelFor

Answer

Using a Python if statement inside the component

Answer

dsl.Condition

Question 3

A company uses Vertex AI Pipelines for ML training. They want to implement continuous training triggered by new data arrival. Which two Google Cloud services should they use to achieve this? (Choose two.)

Accepted Answer

Cloud Functions. Cloud Functions is correct because it can be triggered directly by Cloud Storage events (e.g., object finalize/create) to start a Vertex AI Pipeline run when new data arrives, enabling event-driven continuous training without manual intervention. Cloud Storage is correct because it serves as the source of new data and its event notifications (via Pub/Sub) are the trigger mechanism that Cloud Functions subscribes to, forming the core event-driven architecture.

Answer

Cloud Scheduler

Answer

Vertex AI Experiments

Answer

Cloud Composer

Question 4

A data science team uses Cloud Composer to orchestrate ML workflows. They need to trigger a Vertex AI pipeline after a BigQuery data load completes, and then run a Dataflow job. Which Airflow operator should they use to launch the Vertex AI pipeline?

Choose an answer to begin — your selection is scored in the full session.

Accepted Answer

VertexAIPipelineJobOperator. Option D, VertexAIPipelineJobOperator, is correct because it is the dedicated Airflow operator for submitting and running a Vertex AI pipeline job from Cloud Composer, which is exactly the first task the team needs to orchestrate. Option E, PythonOperator with if-else, is correct because conditional branching based on the pipeline's output can be implemented by a PythonOperator that reads the pipeline result (e.g., via XCom) and evaluates an if-else to decide whether the BigQuery query should run. Option A, BigQueryExecuteQueryOperator, is not marked correct because although it runs BigQuery queries, the question asks for the Airflow features enabling the conditional workflow, not the query execution operator itself. Option B, task dependencies using >>, is not marked correct because >> only sets static ordering between tasks and cannot express runtime conditional branching. Option C, dsl.If, is not marked correct because it is not a valid Airflow construct for conditional task execution in this context.

Answer

BigQueryExecuteQueryOperator

Answer

Task dependencies using >>

Answer

dsl.If

Question 5

An organization wants to implement continuous delivery for their ML model. After a new model is trained and evaluated, they want to automatically deploy it to a staging endpoint, run validation tests, and if passed, promote to production. Which two components should they include in their delivery pipeline? (Choose two.)

Accepted Answer

Conditional component that checks evaluation metrics and promotes if successful. In a Kubeflow Pipelines-style continuous delivery workflow, option C (ModelDeploymentOp to deploy to staging) is correct because it is the component that actually publishes the trained model artifact to a serving endpoint, which is the required first deployment step before validation tests can run against staging. Option A (Conditional component that checks evaluation metrics and promotes if successful) is correct because the pipeline needs a control-flow gate that evaluates the validation results and only triggers promotion to production when the metrics pass the defined threshold. Option B (Dataflow job to preprocess data) is not part of the deployment/promotion stage; preprocessing belongs to the training data preparation phase, not to post-training delivery. Option D (Importer component to bring in the model) is unnecessary here because the model is produced within the same pipeline and passed as an artifact, so no external import step is needed. Option E (A/B testing component to split traffic) is a production traffic-management technique, not a required pipeline component for staging validation and promotion.

Answer

Dataflow job to preprocess data

Answer

Importer component to bring in the model

Answer

A/B testing component to split traffic

Question 6

An ML engineer manages a Vertex AI Pipeline that trains a model and then runs a batch prediction job. The engineer wants to ensure the batch prediction job only runs after the training job completes successfully, and that the pipeline fails immediately if the training job fails. Which approach should the engineer use?

Accepted Answer

Define the batch prediction task with a dependency on the training task by using the `after` parameter in the pipeline definition.. The `after` parameter in Vertex AI Pipelines defines explicit task dependencies. By listing the training task in the `after` parameter of the batch prediction task, the prediction task will not start until training completes successfully. If training fails, the pipeline stops, preventing the prediction task from running. This approach is standard for sequencing tasks and ensuring failure propagation in Vertex AI Pipelines.

Answer

Use a single custom container component that performs both training and batch prediction sequentially within the same component.

Answer

Set the batch prediction task as a parent task and the training task as a child task in the pipeline graph.

Answer

Configure the batch prediction task to retry on failure and set a maximum retry count of 3.

Question 7

An ML engineer is building a Vertex AI Pipeline that includes a custom training component. The component must read a dataset from BigQuery, train a model, and write the model artifact to Cloud Storage. The engineer wants the pipeline to cache the component's output so that if the pipeline is re-run with the same inputs, the training step is skipped. Which combination of component definition and pipeline settings should the engineer use?

Accepted Answer

Set the component's `enable_caching` parameter to True when defining the pipeline task, and ensure the component's inputs and outputs are deterministic.. Vertex AI Pipelines caches component executions based on the component's inputs, outputs, and code. To enable caching, the engineer must set `enable_caching=True` on the pipeline task. This allows the pipeline to skip re-execution if the same inputs are used and the component definition hasn't changed.

Answer

Use a `CacheDataset` artifact as an input to the component and set the pipeline's `cache_mode` to `FORCE_CACHE`.

Answer

Configure the pipeline to use a `CacheConfig` resource with a TTL, and set the component's `output_artifact` to a `Model` type.

Answer

Add a `cache` parameter to the component's `@component` decorator and set it to `True`.

Question 8

An ML engineer is authoring a custom training component for a Vertex AI Pipeline using the Kubeflow Pipelines (KFP) SDK. The component must be reusable across different environments and must accept a dataset URI, a learning rate, and an output model artifact. The engineer wants to define the component using a lightweight Python function that can be compiled without building a custom container image. Which KFP component definition approach should they use?

Accepted Answer

Define the component using the @component decorator from the kfp.v2.dsl module and specify the base image in the decorator.. The @component decorator from the KFP SDK allows defining a component as a Python function with typed inputs and outputs. When compiled, the SDK generates a component specification that uses a base image you specify. This avoids building a custom container and keeps the component reusable. The other options either require custom container builds, use deprecated utilities, or involve manual YAML authoring, which do not match the lightweight Python function requirement.

Answer

Write a YAML component specification file that defines the component's inputs, outputs, and implementation as a shell command.

Answer

Use the kfp.components.func_to_container_op utility to convert the Python function into a component at runtime.

Answer

Create a Dockerfile for the component, build the image, push it to Artifact Registry, and reference it in the pipeline using the component's image URI.

Question 9

An ML engineer is using Vertex AI Pipelines to orchestrate a training workflow. The pipeline includes a custom training component that runs for several hours. The engineer wants to ensure that if the pipeline is preempted or fails, it can resume from the last successful step without re-running completed steps. Which feature should the engineer enable?

Accepted Answer

Set the pipeline's caching policy to use cached results and re-submit the pipeline after failure.. Caching in Vertex AI Pipelines allows re-submitting a failed pipeline to skip already completed components, effectively resuming from the last successful step. By enabling caching, the pipeline reuses outputs of components whose inputs and code have not changed, avoiding redundant computation. This is the correct mechanism for resuming pipelines after preemption or failure.

Answer

Enable Vertex AI Pipelines' built-in retry policy for the training component.

Answer

Set the pipeline's caching policy to always use cached results.

Answer

Configure the pipeline to use a persistent resource and enable checkpointing in the training code.

Question 10

An ML engineer is authoring a Vertex AI Pipeline using the Kubeflow Pipelines (KFP) SDK. The pipeline includes a custom training component that runs on a single worker and takes about 4 hours. The engineer wants to reduce cost by using a preemptible VM, but the training must complete reliably. What should the engineer do?

Accepted Answer

Set the component's task resource spec to use a preemptible VM and enable automatic retries on the task.. The correct approach is to use a preemptible VM for cost savings and enable automatic retries on the training task. Vertex AI Pipelines supports specifying preemptible VMs in the task's resource spec and allows setting retry policies per task. This ensures that if the VM is preempted, the task will restart automatically, providing reliability without manual intervention. Other options either misapply features or require unnecessary complexity.

Answer

Use a preemptible VM and set the component to checkpoint its state to Cloud Storage every hour.

Answer

Configure the component to run as a Vertex AI CustomJob with a preemptible VM and set the pipeline's retry policy to retry the entire pipeline.

Answer

Use a standard VM but set the pipeline to run in a spot instance pool.

Question 11

A team is using Vertex AI Pipelines to orchestrate a workflow that includes a data preprocessing component, a training component, and a deployment component. They want to ensure that the deployment component only runs if the training component completes successfully. Which mechanism should they use?

Accepted Answer

Define a condition on the deployment component that checks if the training component's output artifact exists.. Vertex AI Pipelines supports conditional execution using dsl.Condition. By checking for the existence of an output artifact from the training component, the pipeline can ensure that the deployment component only runs if training succeeded. This is a robust pattern for gating deployment on training success, as a failed training job will not produce the expected artifact.

Answer

Set the deployment component's retry policy to zero and rely on the pipeline's default behavior to skip it on failure.

Answer

Use the after keyword to specify that the deployment component should run after the training component.

Answer

Use a pipeline parameter to pass the training status to the deployment component and have it check the status at runtime.

Question 12

A team is orchestrating a Vertex AI Pipeline that includes a custom training component. They want to ensure the pipeline is reproducible and that artifacts are cached appropriately when inputs have not changed. They also want to control which pipeline runs reuse cached artifacts. Which two actions should they take? (Choose two.)

Accepted Answer

Use the Vertex AI Pipelines API to disable caching for specific components by setting their caching policy to 'never' when inputs are known to change frequently.. To ensure reproducible pipelines with appropriate caching, enable caching at the pipeline level and make component inputs deterministic and versioned. Additionally, use per-component caching policies to disable caching where inputs change frequently. These two actions together provide control over when cached artifacts are reused. The other options either invent non-existent features, misunderstand cache key computation, or suggest destructive practices.

Answer

Store all training data and models in a single Cloud Storage bucket with uniform permissions to simplify cache key computation.

Answer

Set the pipeline's execution mode to 'reproducible' and specify a unique experiment name for each run.

Answer

Manually delete the pipeline's metadata store entries before each run to force cache invalidation.

Question 13

An ML engineer is building a Vertex AI Pipeline that includes a custom training component. The component must use a specific Python package version that is not included in the default Vertex AI training image. The engineer wants to ensure the component runs with the correct dependencies and minimizes build time. What should the engineer do?

Accepted Answer

Specify the package in a requirements.txt file and pass it to the component using the packages_to_install argument of the @dsl.component decorator.. The packages_to_install argument in the @dsl.component decorator is designed to install additional Python packages at runtime without requiring a custom Docker image. This minimizes build time and is ideal for pure Python dependencies that are not pre-installed. Other methods either involve unnecessary build steps or do not correctly install the package.

Answer

Create a custom Docker image that includes the package, push it to Artifact Registry, and reference it in the component definition using the base_image parameter.

Answer

Use the google-cloud-aiplatform package's CustomTrainingJob to run the training with a custom container, and then wrap it in a pipeline component.

Answer

Add the package to the pipeline's runtime configuration by specifying it in the pipeline job's environment variables.

Question 14

A company uses Vertex AI Pipelines to orchestrate a training workflow. They need to ensure that a data preprocessing component runs only once per day, even if the pipeline is triggered multiple times within the same day. The preprocessing component is expensive and should be cached across pipeline runs. What should the ML engineer do?

Accepted Answer

Set the pipeline's caching policy to use a custom cache key that includes the date, and enable caching for the preprocessing component.. To cache a component across pipeline runs but force re-execution daily, the cache key must incorporate a date. Vertex AI Pipelines allows specifying a custom cache key, often by passing a date string as an input parameter. This ensures that runs on the same day share the cached result, while a new day generates a new key and triggers execution. This balances cost savings with data freshness.

Answer

Use a Vertex AI Pipeline schedule with a cron expression that runs the pipeline once per day, and disable manual triggers.

Answer

Set the pipeline's caching policy to always cache, and rely on the component's inputs being identical across runs.

Answer

Configure the preprocessing component to write a marker file to Cloud Storage and check for its existence before running.

Question 15

An ML engineer is building a Vertex AI pipeline where a component must read a model artifact from a previous training step and write evaluation metrics to Vertex ML Metadata. The engineer wants the metrics to be easily queryable and comparable across pipeline runs. Which approach should the engineer use to store the metrics?

Accepted Answer

Output the metrics as a `Metrics` artifact using `google.VertexAI.Metrics` and set it as an output of the component.. Using a Metrics artifact as an output of the component ensures that the metrics are automatically logged to Vertex ML Metadata with lineage to the pipeline run. This makes them queryable and comparable across runs through the Vertex AI UI or API. Other methods either require manual logging or do not provide the structured metadata integration needed for easy comparison.

Answer

Embed the metrics in the component's container logs and use Cloud Logging to extract them.

Answer

Use the `aiplatform.log_metrics` function to log the metrics within the component.

Answer

Write the metrics to a file in Cloud Storage and pass the URI as an output parameter.