GADP

Free GADP practice questions

10 questions from our GADP practice tests, each with the correct answer and a written explanation. Read them with the answers shown, or switch to quiz mode and answer first.

Back to the GADP practice exams

Question 1 of 10, Data Preparation and Ingestion

A company must ingest data nightly from an on-premises Oracle database and a Salesforce instance, apply joins and lookups, and load BigQuery. The team prefers prebuilt connectors and a visual designer, and it needs pipeline lineage. Which service should it use?

  1. ABigQuery Data Transfer Service
  2. BServerless for Apache Spark batches
  3. CCloud Data Fusion Correct answer
  4. DDatastream

Answer and explanation: Cloud Data Fusion offers a visual designer with a large library of prebuilt connectors, including databases and SaaS sources, supports joins and lookups, and tracks lineage. Datastream replicates database changes but does not join with Salesforce, the Data Transfer Service loads supported sources without custom joins, and Serverless for Apache Spark (formerly Dataproc Serverless) requires writing Spark code.

Question 2 of 10, Data Preparation and Ingestion

A manufacturer ingests millions of sensor readings per second and needs single-digit millisecond reads of recent readings by device ID and timestamp. Which storage service should it choose?

  1. AFirestore with one document per reading in a device collection
  2. BBigQuery with the readings table partitioned by ingestion day
  3. CCloud SQL for PostgreSQL with an index on device ID and time
  4. DBigtable with a row key that combines device ID and timestamp Correct answer

Answer and explanation: Bigtable handles very high write throughput with low-latency lookups by row key, and a device-plus-timestamp key supports fast range scans per device. BigQuery is optimized for analytical scans rather than millisecond point reads, Cloud SQL cannot absorb millions of writes per second on one instance, and Firestore is a document store aimed at application data rather than massive time-series ingestion.

Question 3 of 10, Data Preparation and Ingestion

A company has 30 TB of files on an on-premises NFS share and a dedicated 10 Gbps Cloud Interconnect with plenty of spare capacity. It wants a managed, resumable transfer to Cloud Storage with integrity checks. Which option should it use?

  1. ATransfer Appliance, because any dataset over 10 TB must ship offline
  2. BDatabase Migration Service configured with the share as a source
  3. CBigQuery Data Transfer Service scheduled against the NFS share path
  4. DStorage Transfer Service for on-premises data using transfer agents Correct answer

Answer and explanation: With 10 Gbps of spare bandwidth, 30 TB can move online in well under two days, and Storage Transfer Service agents provide managed, resumable transfers with checksums. Transfer Appliance makes sense when bandwidth is the bottleneck, not at a fixed size, BigQuery Data Transfer Service loads into BigQuery from supported sources, and Database Migration Service migrates databases rather than file shares.

Question 4 of 10, Data Analysis and Presentation. Choose 2.

A fraud model in BigQuery ML is trained on data where only 0.5% of transactions are fraudulent, and it predicts 'not fraud' for nearly everything. Which two actions help? (Choose two.)

  1. ARemove all fraudulent rows so the model trains on clean data only
  2. BSet auto_class_weights to TRUE when creating the model Correct answer
  3. CEvaluate with precision, recall, and ROC AUC instead of accuracy Correct answer
  4. DRaise the classification threshold so fewer rows are flagged
  5. EUse ARIMA_PLUS so the model learns fraud trends over time

Answer and explanation: Class weighting makes errors on the rare class count more during training, and metrics such as precision, recall, and ROC AUC reveal performance on the minority class that accuracy hides. Removing fraud rows eliminates the class the model must learn, ARIMA_PLUS forecasts time series rather than classifying transactions, and raising the threshold flags even fewer frauds.

Question 5 of 10, Data Analysis and Presentation

A support team wants to summarize thousands of customer comments stored in a BigQuery table using a Google Gemini model, without moving the data out of BigQuery or writing application code. What should the data practitioner do?

  1. AExport the comments to Cloud Storage and run a Dataflow pipeline that sends each comment to the Gemini API for summarization
  2. BTrain a BigQuery ML boosted tree classifier on the comments and use ML.PREDICT to output a summary per row
  3. CCreate a remote model over a Vertex AI Gemini endpoint through a BigQuery connection and call ML.GENERATE_TEXT Correct answer
  4. DUse the Natural Language API sentiment analysis from a Cloud Run function triggered on each new comment row

Answer and explanation: BigQuery ML remote models connect to Vertex AI Gemini endpoints through a Cloud resource connection, and ML.GENERATE_TEXT runs prompts over table rows using SQL. A boosted tree classifier predicts labels and cannot write summaries, a Dataflow export pipeline moves data and requires code, and sentiment analysis scores tone rather than producing summaries.

Question 6 of 10, Data Analysis and Presentation

A model trained in BigQuery ML shows 98% accuracy on its training data but only 71% on the evaluation data. What is the MOST likely cause, and what should the practitioner do first?

  1. AThe model is underfitting; add many more input features and iterations
  2. BThe model is overfitting; add regularization or simplify the features Correct answer
  3. CThe evaluation data is mislabeled; train again using the evaluation rows
  4. DThe split is random; switch to a fixed split with no evaluation set

Answer and explanation: A large gap between training and evaluation performance is the classic sign of overfitting, which regularization or simpler features help address. Underfitting shows poor performance on both sets, retraining on the evaluation rows would remove the ability to measure generalization, and removing the evaluation set hides the problem rather than fixing it.

Question 7 of 10, Data Pipeline Orchestration

A team needs to chain a few serverless steps: call an external REST API, start a BigQuery job, wait for it, and call a Cloud Run service with the result. They want a lightweight, pay-per-use orchestrator with no environment to manage. Which service should they choose?

  1. AWorkflows Correct answer
  2. BDataproc workflow templates
  3. CCloud Composer
  4. DDataflow

Answer and explanation: Workflows is a serverless orchestrator for sequencing HTTP-based service calls, including Google Cloud APIs, with pay-per-step pricing and no environment. Cloud Composer runs a persistent Airflow environment that is heavier than needed, Dataproc workflow templates orchestrate jobs on a Dataproc cluster, and Dataflow is a data processing engine rather than an API orchestrator.

Question 8 of 10, Data Pipeline Orchestration

A team runs a sequence of dependent Spark jobs that should create a managed Dataproc cluster, run the jobs in order, and delete the cluster afterward, all defined in one reusable YAML resource. What should they use?

  1. AA scheduled query that calls the Spark jobs
  2. BA Dataproc workflow template with a managed cluster Correct answer
  3. CA Cloud Scheduler job that submits each Spark job directly
  4. DAn Eventarc trigger on the Dataproc cluster creation

Answer and explanation: Dataproc workflow templates define a DAG of jobs and can create an ephemeral managed cluster, run the jobs in order, and delete it when done. Cloud Scheduler alone does not manage dependencies or cluster lifecycle, scheduled queries run SQL rather than orchestrating Spark jobs on a cluster, and an Eventarc trigger reacts to events but does not define a job sequence.

Question 9 of 10, Data Management

A compliance team keeps audit logs in Cloud Storage. The logs are rarely read after the first month but must be retained for seven years at the lowest storage cost. Which configuration is MOST cost-effective?

  1. AA lifecycle rule that moves objects to Archive storage after 30 days Correct answer
  2. BA lifecycle rule that moves objects to Nearline storage after 30 days
  3. CEnabling Autoclass so that objects are moved to Coldline after 30 days
  4. DKeeping objects in Standard storage with versioning enabled

Answer and explanation: Archive storage has the lowest at-rest price and suits data kept for years and almost never read, so a SetStorageClass rule to Archive after 30 days is cheapest. Nearline costs more per GB and targets roughly monthly access, Standard with versioning is the most expensive option, and Autoclass moves objects based on access patterns, reaching Coldline only after 90 days without access and Archive only after a year, so it is slower to reach the cheapest tier.

Question 10 of 10, Data Management

A company stores critical data in a dual-region Cloud Storage bucket and needs a guaranteed recovery point objective of 15 minutes for newly written objects if one region fails. What should it enable?

  1. AA second bucket in a different multi-region
  2. BAutoclass with Standard as the starting class
  3. CObject versioning plus soft delete
  4. DTurbo replication on the dual-region bucket Correct answer

Answer and explanation: Turbo replication provides a 15-minute RPO for objects written to dual-region buckets, compared with default replication that targets most objects within an hour and all within 12 hours. Versioning and soft delete protect against deletion, not regional loss, a second bucket needs a copy process with no RPO guarantee, and Autoclass changes storage classes rather than replication speed.