PDE Multiple Select — 40 Questions

Question 1

Which TWO steps are required to deploy a custom scikit-learn model to Vertex AI for online predictions?

Accepted Answer

Save the model using joblib or pickle. Scikit-learn models must be serialized using joblib or pickle to be saved as a model artifact that can be uploaded to Vertex AI. Vertex AI's pre-built prediction containers for scikit-learn expect the model file to be in this format (typically model.joblib or model.pkl) to serve online predictions.

Answer

Write a custom prediction routine

Answer

Containerize the model using Docker

Answer

Create a Vertex AI Endpoint manually

Question 2

A data warehouse team uses Cloud BigQuery for analytics. They want to optimize query performance and reduce costs. Which three actions should they take? (Choose 3)

Accepted Answer

Use partitioned tables on time columns. Partitioning tables on time columns (e.g., DATE, TIMESTAMP) in BigQuery allows the query engine to perform partition pruning, scanning only the relevant partitions instead of the entire table. This directly reduces the amount of data read, lowering query costs and improving performance by limiting I/O to the necessary time range.

Answer

Use automatic reclustering

Answer

Use BI Engine for all queries

Question 3

A company uses Cloud Dataproc for ephemeral clusters to run batch jobs. They want to ensure job reliability and data quality. Which two configuration options should they use? (Choose two.)

Accepted Answer

Use initialization actions for cluster setup.. Option B is correct because initialization actions let you run scripts on every node at cluster creation, so you can install monitoring/validation agents, configure logging, and enforce consistent setup that supports job reliability and data quality across ephemeral clusters. Option E is correct because graceful decommissioning (set via yarn:yarn.resourcemanager.decommissioning.timeout or Dataproc's gracefulDecommissionTimeout) lets running tasks finish before workers are removed, preventing partial job failures and data corruption when nodes are scaled down or preempted. Option A is not appropriate here because preemptible VMs reduce cost but can be reclaimed at any time, which undermines reliability rather than improving it. Option C is not appropriate because idle timeout only deletes idle clusters to save cost; it does not improve job reliability or data quality. Option D is not appropriate because custom machine types tune CPU/memory performance but do not by themselves ensure reliable job completion or data quality.

Answer

Enable preemptible VMs for cost savings.

Answer

Enable idle timeout to automatically delete clusters.

Answer

Use custom machine types for better performance.

Question 4

You are designing a data processing architecture on Google Cloud. You need to ingest data from multiple sources, including streaming events and batch files, and process them to produce a unified dataset for analytics. The solution must support both real-time and historical processing with the same codebase, and be able to handle late-arriving data. Which two Google Cloud services should you use together to achieve this? (Choose two.)

Accepted Answer

Cloud Pub/Sub. To achieve unified batch and stream processing with the same codebase and handle late data, you need a processing engine that supports Apache Beam, such as Cloud Dataflow, and a scalable ingestion service for streaming data, such as Cloud Pub/Sub. Dataflow's Beam model allows you to write one pipeline that works for both batch and streaming, with built-in support for event-time windowing and late data. Pub/Sub provides reliable, scalable ingestion for streaming events.

Answer

Cloud Composer

Answer

Cloud Dataproc

Answer

BigQuery

Question 5

A company is planning to migrate a legacy batch ETL pipeline to Google Cloud. The pipeline involves reading from a relational database, transforming data, and writing to a data warehouse. Which three Google Cloud services can be used as the orchestration layer? (Choose three.)

Accepted Answer

Cloud Scheduler. Cloud Scheduler is a fully managed cron job service that can trigger orchestration workflows on a schedule. It is correct because it can initiate batch ETL pipelines by sending HTTP requests to Cloud Run, Cloud Functions, or Pub/Sub, making it a lightweight orchestration trigger for scheduled batch jobs.

Answer

Cloud Dataproc

Answer

Cloud Dataflow

Question 6

A data scientist needs to perform feature engineering for a machine learning model using Vertex AI. They want to preprocess data using a pipeline that includes scaling, one-hot encoding, and handling missing values. Which TWO services can they use to define and execute this preprocessing pipeline? (Choose 2.)

Accepted Answer

Vertex AI Pipelines. Vertex AI Pipelines is the recommended service for building and running ML pipelines, including preprocessing steps. Alternatively, you can use BigQuery SQL for feature engineering directly on the data, then export the processed data for training. Cloud Dataflow is an option for batch/streaming data processing but is not specific to ML pipelines. Cloud Functions and Dataproc are less suitable for this purpose.

Answer

Cloud Dataproc

Answer

Cloud Dataflow

Answer

Cloud Functions

Question 7

An organization is using BigQuery for analytics. They have a table that is 500 GB and is frequently queried by 'date' and 'region'. They want to optimize query performance and reduce costs. Which TWO actions should they take?

Accepted Answer

Cluster the table by region. Option E is correct because partitioning the table by date means BigQuery only scans the partitions that match the query's date filter, drastically reducing bytes processed and cost for a 500 GB table frequently filtered by date. Option D is correct because clustering the table by region physically sorts data within each partition by region, so queries filtering on region benefit from block pruning and scan less data. Together, partitioning by date and clustering by region match the two most common query predicates and are the standard BigQuery optimization pattern. Option A (authorized view) only controls access to data and does not improve query performance or reduce scan costs. Option B (wildcard table) is used to query multiple similarly named tables with a UNION-like syntax and does not optimize a single table's scans. Option C (materialized views) can help some workloads but is not the primary optimization for a frequently filtered base table and adds storage and refresh overhead, so it is not one of the two best actions here.

Answer

Use an authorized view

Answer

Use a wildcard table

Answer

Use materialized views

Question 8

A data engineer is designing a batch processing system using Cloud Dataproc. Which TWO practices improve performance and reduce costs? (Choose TWO.)

Accepted Answer

Set autoscaling policies based on YARN memory.. Autoscaling policies based on YARN memory allow the cluster to dynamically add or remove worker nodes in response to actual resource demand from running jobs. This prevents over-provisioning (reducing costs) and ensures sufficient resources for job completion (improving performance), as Cloud Dataproc directly monitors YARN memory metrics to trigger scaling actions.

Answer

Always use persistent disks for all nodes.

Answer

Store intermediate data in HDFS.

Answer

Use the largest machine types for master nodes.

Question 9

A company uses Workflows to orchestrate a multi-step data pipeline. One step calls an HTTP endpoint that may take up to 10 minutes, but the default Workflows timeout is too short. They also need to handle transient errors with retries. Which TWO configurations should they apply? (Choose 2)

Accepted Answer

Set a step timeout of 600 seconds for the HTTP call step. Option A is correct because Workflows steps support a per-step timeout, and setting it to 600 seconds (10 minutes) ensures the HTTP call step is allowed to run for the full duration instead of being cut off by the shorter default step timeout. Option E is correct because transient errors (such as 5xx responses or connection resets) should be handled by attaching a retry policy to the step with conditions that match those transient failures, so the step is retried automatically. Option B is not appropriate because Workflows does not use dead letter queues for failed steps; failures are surfaced through execution errors and logs. Option C is wrong because relying on the default retry policy does not let them target transient errors with appropriate conditions, and the default may not retry at all. Option D is wrong because setting the workflow execution timeout to 600 seconds would cap the entire workflow at 10 minutes, which is not the same as giving the HTTP step enough time and could still cause premature termination.

Answer

Configure a dead letter queue for failed steps

Answer

Use the default retry policy on the step

Answer

Set the workflow execution timeout to 600 seconds

Question 10

Which TWO are benefits of using Vertex AI Endpoints for model serving?

Accepted Answer

Integrated monitoring for prediction latency and error rates.. Option B is correct because Vertex AI Endpoints provide integrated monitoring that surfaces prediction latency and error rates, letting you track online serving health and performance directly in the console. Option C is correct because Vertex AI Endpoints support automatic scaling (autoscaling) based on traffic, adjusting the number of deployed model replicas to match request load. Options A, D, and E are not benefits of Endpoints: batch prediction is a separate Vertex AI feature (Batch Prediction jobs) rather than an Endpoint capability, automatic retraining on drift is not a built-in Endpoint function, and A/B testing requires explicit traffic-split configuration rather than working with no additional setup.

Answer

Batch prediction support out of the box.

Answer

Automatic model retraining when drift is detected.

Answer

Built-in support for A/B testing without any additional configuration.

Question 11

A company uses Cloud Dataproc for large-scale Spark jobs. They notice that some jobs are failing due to insufficient memory on the worker nodes. They want to improve memory management without over-provisioning. Which three configurations should they apply? (Choose 3)

Accepted Answer

Set spark.executor.memory to a value that fits within the node memory. Setting spark.executor.memory to a value that fits within the node memory ensures that each executor does not exceed the available RAM on a worker node, preventing out-of-memory (OOM) errors. This configuration directly controls the heap size allocated to each executor, and when combined with spark.executor.cores and spark.executor.instances, it allows precise memory budgeting per node. Over-provisioning is avoided by calculating the maximum safe executor memory as (node memory - OS overhead - HDFS cache) / number of executors per node.

Answer

Use local SSDs for temporary storage

Answer

Use preemptible worker nodes for volatile tasks

Question 12

A company uses BigQuery for analytics on petabyte-scale data. They want to improve query performance by denormalizing schemas and reducing joins. Which TWO BigQuery features should they use? (Choose 2)

Accepted Answer

Clustering on frequently filtered columns. Option E is correct because BigQuery's native support for nested and repeated fields (ARRAY>) lets you store parent-child relationships inside a single table, which is the standard way to denormalize schemas and eliminate joins while preserving relational structure. Option A is correct because clustering on frequently filtered columns physically co-locates related rows within partitions, so BigQuery scans far less data and returns results faster for filtered queries on the denormalized table. Option B is not a BigQuery feature for denormalization; subqueries still execute as separate query blocks and do not remove join costs. Option C is wrong because external tables read data from Cloud Storage and generally offer worse performance than native tables, not better denormalized performance. Option D is wrong because partitioning by date improves pruning on time filters but does not denormalize schemas or reduce joins.

Answer

Using subqueries instead of JOINs

Answer

External tables reading from Cloud Storage

Answer

Table partitioning by date

Question 13

Which TWO are common causes of prediction bias in a deployed machine learning model in production?

Accepted Answer

Data drift between training and serving data distributions.. Option B is correct because data drift occurs when the statistical distribution of the input features or target changes between the training environment and the live serving environment, causing the model's learned mappings to become stale and its predictions to be systematically biased. Option E is correct because training-serving skew arises when feature engineering logic differs between the training pipeline and the production inference path (for example, different imputation, normalization, or aggregation code), so the model receives inputs at serving time that do not match what it learned from, producing biased predictions. Option A is not a cause of bias — high accuracy is generally desirable and does not by itself indicate or produce prediction bias. Option C is not correct in this context because overfitting primarily harms generalization and variance rather than being a canonical cause of prediction bias in production, and it is a training-time issue rather than the drift/skew mechanisms described. Option D is not correct because low latency is a performance characteristic of the serving system and has no direct causal relationship with prediction bias.

Answer

Model accuracy is too high.

Answer

Model is overfitted to training data.

Answer

Low latency predictions.

Question 14

Your company uses Cloud Composer to orchestrate a data pipeline that includes Dataproc Spark jobs and BigQuery load operations. You need to pass the output file path from the Spark job to the next BigQuery task in the DAG. Which two mechanisms can you use to share data between tasks? (Choose TWO.)

Accepted Answer

Write the output path to a Cloud Storage object and read it in the next task.. Option E is correct because Airflow XComs are the native, built-in mechanism for passing small pieces of data such as a file path between tasks in a DAG: the Spark task calls xcom_push (or returns a value) and the BigQuery task uses xcom_pull with the task_id/key to retrieve it. Option C is correct because writing the output path to a Cloud Storage object and reading it in the next task is a durable, decoupled way to share data between tasks, especially when the value is larger or must persist beyond the DAG run; Cloud Composer tasks can use the GCS operators/sensors or the google-cloud-storage client to write and read the object. Option A is not appropriate because Cloud Composer (Airflow) Variables are global key-value configuration stores, not per-run task-to-task data channels, so using them for a run-specific output path risks race conditions and stale values. Option B is not appropriate because Pub/Sub is an asynchronous messaging service, not a synchronous task-to-task handoff mechanism, and it adds unnecessary complexity and latency for passing a single path within one DAG run. Option D is not appropriate because using BigQuery as an intermediary to store a transient output path is heavyweight and semantically wrong; BigQuery is for analytical data, not for orchestrating task metadata.

Answer

Store the output path as a Cloud Composer variable.

Answer

Publish the output path to a Pub/Sub topic and subscribe in the next task.

Answer

Use BigQuery as an intermediary to store the output path.

Question 15

Which TWO actions are recommended to improve the reliability of a Cloud Dataflow streaming pipeline that processes event data from Pub/Sub?

Accepted Answer

Enable Dataflow Streaming Engine.. Option B is correct because Dataflow Streaming Engine moves pipeline state and shuffle execution out of the worker VMs and into the Dataflow service, which improves reliability by decoupling state management from worker lifecycle events such as restarts, autoscaling, and updates. Option C is correct because enabling exactly-once processing sinks (for example, BigQuery with row-level deduplication via insertion IDs) prevents duplicate or lost records when retries occur, which is essential for reliable event processing from Pub/Sub. Option A is not recommended because a 10-second acknowledgment deadline is too short for many streaming workloads and would cause premature redelivery and duplicate processing rather than improving reliability. Option D is incorrect because disabling autoscaling does not improve reliability and can actually reduce throughput and resilience during load spikes. Option E is incorrect because micro-batching with a small batch size is not a Dataflow reliability best practice and can increase overhead and latency without addressing state or delivery guarantees.

Answer

Use a pull subscription with a 10-second acknowledgment deadline.

Answer

Disable autoscaling to prevent worker churn.

Answer

Use micro-batch processing with a small batch size.

Question 16

A company uses Cloud Logging to monitor application errors. They want to set up real-time notifications for critical errors. Which two actions are essential? (Choose two.)

Accepted Answer

Create a log-based metric for critical errors.. Option A is correct because a log-based metric is the mechanism that converts matching log entries (e.g., a filter for severity=ERROR or specific critical error text) into a numeric time series that Cloud Monitoring can evaluate; without it there is no metric for alerting. Option E is correct because a Cloud Monitoring alert policy is what actually triggers real-time notifications, and it must be based on the log-based metric created in A to fire when critical errors occur. Option B is not essential since exporting to BigQuery is for retrospective analysis, not real-time alerting. Option C is not essential because a Pub/Sub notification on a log sink is a separate log-routing feature and does not by itself create a Monitoring alert with notification channels. Option D is not essential because VPC Flow Logs capture network traffic metadata and are unrelated to application error alerting.

Answer

Export logs to BigQuery for later analysis.

Answer

Create a Cloud Pub/Sub notification directly on the log sink.

Answer

Enable VPC Flow Logs to capture network traffic.

Question 17

A data engineering team is operationalizing a machine learning model for real-time fraud detection. The model must process transactions with sub-100ms latency and be highly available. Which TWO strategies should the team implement?

Accepted Answer

Deploy the model to multiple Google Cloud regions for failover.. Option A is correct because deploying the model to multiple Google Cloud regions provides geographic redundancy and failover, which directly supports the high-availability requirement for a real-time fraud detection service. Option D is correct because pruning or quantizing the model reduces its size and computational cost, which helps achieve the sub-100ms latency target for real-time inference. Option B is incorrect because a single zone creates a single point of failure and does not meet the high-availability requirement, even if it reduces cross-zone latency. Option C is incorrect because Cloud Batch is designed for asynchronous, batch-oriented workloads, not real-time sub-100ms predictions. Option E is incorrect because loading the model from Cloud Storage on every request adds significant network and I/O latency, making the sub-100ms target impractical.

Answer

Deploy the model to a single zone to minimize cross-zone latency.

Answer

Use Cloud Batch for asynchronous prediction.

Answer

Store the model in Cloud Storage and load it on each request.

Question 18

Your team is running a Dataflow streaming pipeline that reads from Pub/Sub, transforms data, and writes to BigQuery. You notice that the pipeline's backlog is growing and the processing latency has increased from seconds to minutes. You need to diagnose and resolve the issue. Which TWO actions should you take? (Choose two.)

Accepted Answer

Increase the maximum number of workers in the pipeline's autoscaling configuration to allow more compute resources.. Option E is correct because the Dataflow monitoring dashboard exposes streaming-specific metrics such as system lag, data freshness, and worker throughput, which are exactly the signals needed to determine whether the growing backlog is caused by insufficient compute, a hot key, or a slow sink. Option D is correct because a streaming pipeline with a growing backlog and rising latency is typically under-provisioned, and raising the maximum number of workers in the autoscaling configuration lets Dataflow horizontally scale the worker pool to absorb the increased load. Option A is not appropriate because stopping and restarting the pipeline interrupts streaming state and is unnecessary when autoscaling can be adjusted on a running job. Option B is not appropriate because increasing the WriteToBigQuery batch size does not address a Pub/Sub backlog and can actually increase latency and memory pressure. Option C is not appropriate because a dead-letter queue for failed messages does not relieve the reprocessing load of successfully processed but backlogged messages.

Answer

Stop the pipeline, increase the number of workers in the streaming engine configuration, and restart it.

Answer

Increase the batch size in the WriteToBigQuery transform to reduce I/O operations.

Answer

Configure a dead-letter queue in Cloud Storage for failed messages to reduce reprocessing load.

Question 19

Which THREE steps are required to set up a continuous training pipeline on Google Cloud using Vertex AI?

Accepted Answer

Create a Vertex AI Pipeline to orchestrate data preprocessing, training, and model evaluation.. Option B is correct because a Vertex AI Pipeline is the native orchestration mechanism that chains the preprocessing, training, and evaluation components into a repeatable, versioned workflow, which is the backbone of a continuous training pipeline. Option C is correct because continuous training requires an automated trigger — Cloud Scheduler for time-based runs or Cloud Build (often via Eventarc/pub-sub on new data) — to invoke the pipeline without manual intervention. Option E is correct because continuous training must include automated model evaluation and promotion logic, such as comparing accuracy against a threshold and deploying to a Vertex AI endpoint only when the gate passes, enabling CI/CD-style model rollout. Option A is not appropriate because a single Compute Engine VM with cron is a manual, non-managed approach that bypasses Vertex AI's managed pipelines, scaling, and metadata tracking. Option D is not appropriate because manually uploading models to the Vertex AI Model Registry after each run defeats automation; the pipeline should register models programmatically as a step.

Answer

Run training on a single Compute Engine VM with a cron job.

Answer

Manually upload the model to Vertex AI Model Registry after each training run.

Question 20

A company is designing a CI/CD pipeline for their ML models using Cloud Build and Vertex AI. Which TWO practices should they adopt to ensure reliable and reproducible deployments?

Accepted Answer

Use immutable container images with version tags for each model deployment. Option C is correct because using immutable container images with version tags ensures that each model deployment is tied to a specific, unchangeable artifact, which is essential for reproducibility and rollback in a CI/CD pipeline. Option D is correct because including unit tests for data preprocessing and feature engineering code in the pipeline catches errors early and verifies that the transformation logic behaves consistently, which directly supports reliable and reproducible ML deployments. Option A is not required for reproducibility and would slow the pipeline; manual approval is a governance choice, not a technical practice for reliable, reproducible deployment. Option B is wrong because storing artifacts in a single bucket without versioning prevents tracking and reproducing specific model versions. Option E is wrong because deploying every model version directly to production bypasses validation and testing, undermining reliability and reproducibility.

Answer

Require manual approval for every model change before deployment

Answer

Store all model artifacts in a single Cloud Storage bucket without versioning

Answer

Deploy every model version directly to production for immediate use