GCP-ADP Practice Exam 4 — 20 Questions

Question 1

You are designing a Dataflow pipeline that reads from Cloud Storage and writes to BigQuery. The input files are in Avro format and contain a field with a timestamp that is often null. You need to ensure that records with null timestamps are not dropped but are assigned a default timestamp of the pipeline start time. Which approach should you use?

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

Accepted Answer

Use a ParDo transform to check for null timestamps and replace them with the pipeline start time using a side input.. To replace null timestamps with the pipeline start time, a ParDo transform with a side input is appropriate. The side input can be a singleton PCollection containing the start time, which is computed once and available to all workers. This ensures each record with a null timestamp is assigned the default without dropping data.

Answer

Use a Combine transform to calculate the minimum timestamp and replace nulls with that value.

Answer

Use a GroupByKey transform to group records by timestamp and filter out nulls.

Answer

Use a CoGroupByKey transform to join records with a default timestamp PCollection.

Question 2

Which TWO are valid methods to trigger a Dataflow job?

Accepted Answer

Cloud Console. Option A (Cloud Console) is correct because Google Cloud Dataflow jobs can be created and launched directly from the Cloud Console UI, where you select a template or pipeline and configure parameters before clicking Run. Option C (Cloud Scheduler) is correct because Cloud Scheduler can send an HTTP request to the Dataflow REST API (or trigger a Cloud Function that calls it) to start a Dataflow job on a cron schedule, making it a valid programmatic trigger method. Option B (Direct keyboard input) is not a valid trigger mechanism for Dataflow jobs, as Dataflow has no such input channel. Option D (GCS bucket browsing) is not a trigger method; browsing objects in Cloud Storage does not launch a Dataflow job. Option E (Cloud DNS lookup) is unrelated to Dataflow, as Cloud DNS resolves domain names and cannot start data processing jobs.

Answer

Direct keyboard input

Answer

GCS bucket browsing

Answer

Cloud DNS lookup

Question 3

You are using Dataflow to read from Pub/Sub. The pipeline is failing due to malformed messages. How can you handle these without crashing?

Accepted Answer

Use a Dead Letter Queue (DLQ). A Dead Letter Queue pattern in Dataflow captures messages that fail transformation or parsing so the pipeline can continue processing valid records. In Apache Beam you implement this by branching the pipeline: successful records go to the main sink, and failed records (caught via try/except or a tagged output) are written to a separate Pub/Sub topic or BigQuery table for later inspection. This prevents a single malformed message from killing the streaming job.

Answer

Ignore errors

Answer

Increase memory

Answer

Restart the pipeline

Question 4

You are cleaning data using Dataprep. You want to save the final dataset in BigQuery. What do you do?

Accepted Answer

Use the Export feature. In Dataprep (now Dataprep by Trifacta), after cleaning and transforming a dataset in the flow, you publish results using the Export (or 'Run job'/output) feature, where you specify BigQuery as the output destination and configure the target dataset and table. This writes the transformed data into BigQuery. Copying files or running SQL manually is not how Dataprep delivers results.

Answer

Copy the file from GCS

Answer

Use a SQL query

Answer

Wait for Dataflow to finish

Question 5

Which TWO of the following are true about Dataplex assets? (Choose two)

Accepted Answer

Assets inherit the location of the zone. Option B is correct because a Dataplex asset is always created inside a zone, and it automatically inherits that zone's location (region), so you cannot place an asset in a different location from its parent zone. Option C is correct because an asset is a child resource of exactly one zone in the Dataplex hierarchy (lake > zone > asset); it cannot be attached to multiple zones simultaneously. Option A is wrong because an asset must reference an underlying storage resource, such as a Cloud Storage bucket, BigQuery dataset, or BigLake table, so it cannot exist without one. Option D is wrong because assets can represent Cloud Storage buckets/objects, BigQuery datasets/tables, and BigLake tables, not only BigQuery tables. Option E is wrong because Dataplex assets are metadata representations that point to the existing physical data; they do not copy or duplicate the underlying data.

Answer

Assets can be created without a storage resource

Answer

Assets can only be BigQuery tables

Answer

Assets duplicate the physical data

Question 6

You are implementing a security strategy for your data lake. Which THREE of the following are supported ways to control access to data within a Dataplex Lake? (Choose three)

Accepted Answer

Grant Dataplex-specific roles to users. Option A is correct because Dataplex Lake access is governed through IAM by granting Dataplex-specific roles (such as roles/dataplex.viewer, roles/dataplex.editor, or roles/dataplex.admin) to users, groups, or service accounts, which controls who can view or manage the lake and its zones/assets. Option B is correct because Dataplex assets point to underlying storage such as Cloud Storage buckets or BigQuery datasets, and IAM roles applied directly to those underlying resources (for example roles/storage.objectViewer or roles/bigquery.dataViewer) are honored and provide an additional supported layer of data access control. Option C is correct because Dataplex integrates with BigQuery and Data Catalog Policy Tags (via the Dataplex/Data Catalog taxonomy and policy tag framework) to enforce column-level security, restricting access to specific columns of sensitive data. Option D is not correct because IP-based firewalls are a network-level control (e.g., VPC firewall rules) and are not a supported mechanism for controlling access to data within a Dataplex Lake. Option E is not correct because IAM roles are not assigned to individual files; permissions are granted at the resource hierarchy level (project, bucket, dataset, table, or Dataplex entity), not per file.

Answer

Use IP-based firewalls for all datasets

Answer

Assign roles to individual files

Question 7

A colleague must load a 200 GB uncompressed CSV file from a Cloud Storage bucket into a BigQuery table once, with the lowest possible load cost and no transformation. They want the simplest option that BigQuery handles natively. What should they use?

Accepted Answer

bq load with the CSV file specified as a Cloud Storage URI. Loading a file straight from Cloud Storage with the bq load command is BigQuery's native bulk ingest path and does not require a separate compute service, so it is both the simplest and the least expensive option for an untransformed one-time CSV load. Dataflow, Cloud Data Fusion, and Dataproc all add their own compute charges and pipeline complexity that the scenario does not need.

Answer

A Cloud Data Fusion pipeline with a Cloud Storage source and BigQuery sink

Answer

A Dataflow pipeline using the TextIO source and BigQueryIO sink

Answer

A Dataproc job running Spark to read the CSV and write to BigQuery

Question 8

You want to dynamically generate tasks in an Airflow DAG based on the number of files in a GCS bucket. What is the standard way to do this?

Accepted Answer

Use a Python loop inside the DAG file to instantiate tasks.. The standard way to dynamically generate tasks in Airflow based on external conditions (like the number of files in a GCS bucket) is to use a Python loop inside the DAG file. At DAG parse time, the loop runs and creates task instances dynamically based on the current state (e.g., listing files via a GCS hook). This leverages Airflow's core design: DAGs are Python scripts that are executed to build the task graph. The loop can call an operator (e.g., PythonOperator, GCSToBigQueryOperator) for each file, generating tasks with unique task_ids. This pattern is widely documented and recommended for dynamic task generation, as long as the number of tasks is not excessively large and the DAG is idempotent.

Answer

Use the 'SubDagOperator'.

Answer

Configure an external metadata service to inject tasks.

Answer

Use the 'DynamicTaskGroup' operator in the Airflow library.

Question 9

Which GCS storage class is most cost-effective for data accessed only once per year?

Accepted Answer

Archive. Archive is the most cost-effective GCS storage class for data accessed only once per year because it is designed for long-term storage of data that is accessed less than once a year. Archive storage has the lowest storage cost among GCS classes, but it also has the highest retrieval cost and a minimum storage duration of 365 days. For data accessed only once per year, Archive provides the best cost savings.

Answer

Standard

Answer

Coldline

Answer

Nearline

Question 10

Which THREE of the following are necessary for a production-ready Dataflow pipeline?

Accepted Answer

Error handling. Error handling (B) is necessary because a production Dataflow pipeline must gracefully manage failures such as malformed input records, transient network issues, or failed side inputs, typically via try/except blocks, dead-letter queues, or retry policies, so that a single bad element does not halt the entire streaming or batch job. Logging (C) is required to capture runtime events, warnings, and errors through tools like Cloud Logging (Stackdriver), enabling operators to diagnose issues, audit data flow, and trace pipeline behavior across workers. Monitoring (E) is essential to observe pipeline health and performance metrics such as system lag, throughput, watermark progression, and worker CPU/memory in Cloud Monitoring, allowing alerting and timely intervention before SLA breaches. The unmarked options do not belong: desktop-based processing (A) is irrelevant because Dataflow is a fully managed, cloud-based service that runs on Google Cloud infrastructure, not on a local desktop, and manual refreshes (D) contradict the automation and continuous operation expected of a production pipeline, which should run on schedules or streaming triggers without human intervention.

Answer

Desktop-based processing

Answer

Manual refreshes

Question 11

You are using Dataflow with a custom container. Which command do you use to specify the container image?

Accepted Answer

--sdk_container_image. When running an Apache Beam pipeline on Dataflow with a custom container, the pipeline option that specifies the container image is --sdk_container_image. This tells Dataflow to pull the specified image from Artifact Registry or Container Registry and use it as the SDK harness environment for the workers. The other flags do not exist in the Dataflow pipeline options.

Answer

--container_path

Answer

--custom_image

Answer

--worker_image

Question 12

A healthcare organization needs to analyze sensitive patient data in BigQuery while ensuring that analysts can see aggregated results but cannot view individual patient identifiers. Which Google Cloud feature should they implement?

Accepted Answer

Column-level security with policy tags. Column-level security with policy tags allows classifying sensitive columns and granting access only to authorized users. Analysts without the policy tag role can query other columns and compute aggregates but cannot see patient identifiers. Other options either restrict rows, use less scalable view patterns, or address network boundaries rather than column visibility.

Answer

Row-level security with filter predicates

Answer

BigQuery authorized views

Answer

VPC Service Controls perimeter around BigQuery

Question 13

You are creating a Looker Studio report that must allow users to interactively explore sales data. The report should include controls that let users filter data by product category and date range. Which two controls should you add? (Choose two.)

Accepted Answer

A date range control.. A filter control on the product category dimension and a date range control are the standard Looker Studio controls for interactive filtering by category and time. They apply to all charts on the page, enabling users to dynamically explore sales data. The other options either do not provide filtering or are not interactive.

Answer

A parameter control to adjust the date granularity.

Answer

A chart-level filter on each chart for product category.

Answer

A data control to switch between different data sources.

Question 14

Which tool is used to monitor the performance of your Dataflow pipelines?

Accepted Answer

Cloud Monitoring. Cloud Monitoring is the GCP service that collects and visualizes operational metrics from Dataflow pipelines, including system lag, throughput, worker CPU/memory, and autoscaling behavior. Dataflow natively publishes these metrics to Cloud Monitoring, where you can build dashboards and alerting policies. Cloud Logging captures log entries (errors, debug output) rather than performance metrics, so it is not the monitoring tool.

Answer

Cloud Logging

Answer

BigQuery Console

Answer

Dataprep

Question 15

When designing a Looker dashboard, you need to display a value that is the average of a specific metric across the entire dashboard. What is the best method?

Accepted Answer

Single Value visualization.. A Single Value visualization in Looker is designed to display one aggregated number prominently, such as an average, sum, or count, and it can compute that value across the entire dashboard's underlying query. It is the standard tile for KPI-style metrics like an overall average.

Answer

Treemap.

Answer

Data table.

Answer

Waterfall chart.

Question 16

In Cloud Composer, what is the default behavior if a DAG execution takes longer than the 'dagrun_timeout' setting?

Accepted Answer

The DAG run is marked as failed.. In Cloud Composer (Apache Airflow), if a DAG run exceeds the dagrun_timeout setting, the scheduler marks the DAG run as failed. This prevents stuck runs from consuming resources indefinitely and allows alerting or retry logic to kick in.

Answer

The DAG enters a 'running' state indefinitely.

Answer

The DAG run is paused.

Answer

The scheduler automatically restarts the DAG.

Question 17

You want to share a BigQuery dataset with a partner who uses a different Google Cloud project. What is the most secure way to grant them access?

Accepted Answer

Grant the partner's service account the 'BigQuery Data Viewer' role on the dataset.. Granting the partner's service account the BigQuery Data Viewer role on the dataset is the most secure approach because it follows the principle of least privilege: the partner can read data in that specific dataset only, using their own identity, with no credential sharing. IAM roles can be granted at the dataset level, scoping access precisely and allowing easy revocation.

Answer

Make the dataset public.

Answer

Provide the partner with your project's service account key.

Answer

Export the data to a public Cloud Storage bucket.

Question 18

A data team is building a BigQuery ML model to predict customer churn directly from a BigQuery table. They want to evaluate the model's quality and understand which input features influenced its predictions. Which two BigQuery ML capabilities should they use to accomplish these goals? (Choose two.)

Accepted Answer

Use ML.EXPLAIN_PREDICT to return per-row feature attributions alongside each prediction.. ML.EVALUATE measures the churn model's predictive quality with classification metrics on held-out data, while ML.EXPLAIN_PREDICT returns per-row feature attributions that reveal which inputs drove each prediction. Together they cover both evaluation and interpretability. The other functions describe training data, training progress, or generate embeddings, none of which answers how good the model is or why it predicts churn.

Answer

Use ML.TRAINING_INFO to inspect the iteration history of the model's training run.

Answer

Use ML.GENERATE_EMBEDDING to convert customer records into vector embeddings for similarity search.

Answer

Use ML.FEATURE_INFO to list the summary statistics of each feature column used during training.

Question 19

You are identifying trends in a large dataset using BigQuery. Which SQL function is best suited to identify the top 5 products sold per region?

Accepted Answer

RANK() OVER (PARTITION BY region ORDER BY sales DESC).. RANK() OVER (PARTITION BY region ORDER BY sales DESC) assigns a rank to each product within each region based on sales, so filtering for rank <= 5 yields the top 5 products per region. This is the standard SQL window-function pattern for per-group top-N queries.

Answer

GROUP BY region, product.

Answer

SELECT TOP 5 * FROM sales.

Answer

SUM(sales) OVER (ORDER BY region).

Question 20

You manage a BigQuery dataset that contains a table with 500 TB of historical order data. Analysts frequently run queries that filter on the `order_date` column. Query costs are rising because every query scans the entire table. You want to reduce the bytes scanned and cost without changing the table name or existing SQL. What should you do?

Accepted Answer

Partition the table by the `order_date` column so queries filtering on that column scan only relevant partitions.. Partitioning a table by a column that queries commonly filter on lets BigQuery prune partitions and read only the data needed, which reduces bytes scanned and therefore cost. Because partitioning does not change the table name or require SQL changes, it satisfies all the constraints. The other choices either require SQL rewrites, do not affect scan volume, or block queries instead of optimizing them.

Answer

Set a custom query quota on the project so analysts cannot run queries that scan more than 1 TB.

Answer

Enable the BigQuery BI Engine reservation on the dataset to cache query results and lower cost.

Answer

Create a materialized view that pre-aggregates orders by date and instruct analysts to query the view instead.