Google Cloud Professional Data Engineer Practice Questions¶

40 scenario-based questions weighted by exam domain.

Exam Domain Breakdown¶

Cert page: exams/gcp/data-engineer/

Domain 1: Design Data Processing Systems (Questions 1-9)¶

Question 1¶

Scenario: A company needs to build a data pipeline that processes streaming data from IoT sensors in real-time, detects anomalies, and stores results for further analysis. The pipeline should handle late-arriving data and provide exactly-once processing guarantees. What architecture should they design?

A. Pub/Sub for durable ingestion, Dataflow streaming with windowing, watermarks and exactly-once processing for anomaly detection, and BigQuery for storage B. Batch processing with Cloud Scheduler launching hourly jobs over files in Cloud Storage C. Cloud Functions triggered by Cloud Storage object creation, which handle single files only D. Direct writes from the sensors to BigQuery using the streaming API, with anomaly detection run later

Question 2¶

Scenario: A retail company has historical sales data in BigQuery and needs to forecast future sales. They have data scientists who prefer SQL and don't want to manage ML infrastructure. What should they use?

A. Agent Platform AutoML (formerly Vertex AI AutoML) forecasting, which is managed but is driven from the console or Python SDK rather than SQL B. Custom ML model on Compute Engine using scikit-learn, which gives full control but requires managing VMs, environments, and serving C. An external ML platform fed by an export job D. BigQuery ML with time series forecasting (ARIMA_PLUS), all in SQL

Question 3¶

Scenario: A company needs to migrate their on-premises Hadoop cluster to GCP. They want to minimize code changes and continue using existing Spark jobs. They also want to reduce costs by only paying when jobs run. What should they use?

A. Persistent Dataproc cluster running around the clock with the Spark jobs, which keeps code unchanged but charges for idle hours B. Run Hadoop on Compute Engine VMs managed by the team, which preserves the code but needs patching, scaling, and paying for idle VMs C. Rewrite everything for Dataflow using the Beam SDK D. Dataproc with ephemeral clusters created per job and Cloud Storage for data

Question 4¶

Scenario: A data team needs to design a data lake architecture that supports both batch and streaming data, enables schema evolution, and allows both SQL and programmatic access. What storage format and tools should they use?

A. CSV files in Cloud Storage with a header row, which have no schema evolution B. Parquet/Delta Lake format in Cloud Storage with BigQuery external tables and Dataproc for processing C. JSON in Cloud SQL tables, which is not designed for lake-scale batch and streaming data D. Avro in Bigtable, which stores row-oriented records in a wide-column NoSQL database and does not serve SQL analytics over lake files

Question 5¶

Scenario: A company needs to process clickstream data that arrives at 1 million events per second. They need to aggregate data by user session and detect patterns. Processing should be cost-effective for continuous workloads. What should they use?

A. Cloud Run services that receive each event over HTTP, which scale quickly but have no windowing or event-time state B. Dataflow with streaming pipeline using session windows C. BigQuery streaming inserts with SQL sessionization after load, which ingests the events but does not detect patterns in flight D. Cloud Functions triggered by Pub/Sub for each event, which hit concurrency limits at 1 million events per second and keep no session state

Question 6¶

Scenario: A financial company needs to analyze trading data with sub-second query latency. Data is time-series with high write throughput (millions of rows per second). Queries filter by time range and symbol. What database should they use?

A. Cloud SQL with a composite index on symbol and time, which does not scale to millions of writes per second B. BigQuery with partitioned and clustered tables, which typically returns in seconds, not sub-second C. Firestore with a collection per symbol, which is not designed for this write rate D. Cloud Bigtable with a symbol#timestamp row key design, which gives single-digit millisecond latency at very high write throughput and efficient range scans

Question 7¶

Scenario: A company needs to design a real-time recommendation engine. User interactions should update recommendations within seconds. The system should handle 10,000 recommendations per second with sub-100ms latency. What architecture should they design?

A. Cloud SQL with a read replica and an in-memory cache in front, which improves read latency but cannot absorb the update rate or scale out B. A batch ML model retrained daily and loaded into the serving layer overnight, which cannot reflect interactions within seconds C. BigQuery for real-time queries over the interaction table, which scales for analytics but has latency well over 100 ms per query D. Recommendations API backed by Agent Platform Feature Store, Bigtable for feature serving, and streaming updates via Dataflow

Question 8¶

Scenario: A company has data in multiple sources: Cloud SQL for MySQL databases, Spanner, and CSV files in Cloud Storage. They need a unified view for analytics without maintaining complex ETL pipelines. What should they implement?

A. Use views in each source system, which expose data separately B. Build custom ETL scripts for each source on a schedule C. Copy all data to BigQuery with scheduled export jobs for each source, which duplicates data and needs ongoing pipeline maintenance D. BigQuery federated queries for the databases, external tables for the CSV files, and Data Fusion for transformations

Question 9¶

Scenario: A healthcare company needs to ensure their data processing systems comply with HIPAA. They need encryption at rest and in transit, access audit logs, and data lineage tracking. What should they implement?

A. No special requirements, since Google Cloud services are already certified for HIPAA and nothing extra needs to be configured by the customer B. Default settings, which encrypt data with Google-managed keys but give no perimeter controls, customer key control, or lineage C. CMEK for encryption, VPC Service Controls for perimeter security, Cloud Audit Logs, and Dataplex data lineage D. Self-managed encryption keys held in an on-premises HSM and added to each pipeline in code, which is complex and does not give lineage or perimeter security

Domain 2: Ingest and Process Data (Questions 10-19)¶

Question 10¶

Scenario: A company needs to ingest data from Apache Kafka on-premises into GCP for processing. They want minimal changes to existing Kafka producers. What ingestion method should they use?

A. Direct writes to BigQuery from the Kafka producers using the streaming API, which requires changing every producer B. Rewrite producers for Pub/Sub using the client libraries, which means application code changes in every Kafka producer C. Pub/Sub with Kafka connector or Confluent Cloud integration D. File transfer of log segments to Cloud Storage on a schedule, which is batch and does not carry the live topic stream

Question 11¶

Scenario: A company receives daily files from partners via SFTP. Files need to be validated, transformed, and loaded into BigQuery. Invalid files should be quarantined for review. What pipeline should they build?

A. Compute Engine polling the SFTP folder from a VM on a timer, which needs a server to manage B. Cloud Functions triggered by Cloud Storage object creation to validate each file, Dataflow for transformation, and a BigQuery load, with invalid files moved to a quarantine bucket C. Manual processing by an analyst who checks each file and loads it by hand D. Direct BigQuery load without validation of the file contents, which gives no quarantine

Question 12¶

Scenario: A Dataflow streaming job processes messages from Pub/Sub. Some messages are malformed and cause processing errors. Instead of failing the pipeline, these messages should be handled separately. What should they implement?

A. Ignore errors by catching exceptions and skipping the records, which loses the malformed messages B. Stop processing on the first error so that an engineer can inspect the message C. Dead-letter pattern: catch errors in a DoFn, route bad messages to a side output such as a Pub/Sub topic or BigQuery table, and keep valid traffic flowing D. Let the pipeline fail and rely on Dataflow's automatic retries, which retry forever on bad input

Question 13¶

Scenario: A company needs to process change data capture (CDC) events from their MySQL database and replicate changes to BigQuery in near real-time. What should they use?

A. mysqldump to Cloud Storage on a schedule, which captures a snapshot at one moment and so cannot replicate changes in near real-time B. Datastream for CDC capture, Dataflow for transformation, BigQuery as destination C. Manual queries run by an analyst against the source tables to find changes, which cannot keep up with a busy database D. Daily full table exports loaded into BigQuery, which copy unchanged rows each time and delay changes by up to a day

Question 14¶

Scenario: A Dataflow job processing large datasets runs slowly. Analysis shows data skew - most data goes to a few workers while others are idle. How should they fix this?

A. Implement a Combine.perKey with fanout or use Reshuffle to rebalance data B. Use smaller machine types for the workers C. Add more workers to the job, which gives more capacity but sends the same hot keys to the same few workers D. Reduce parallelism so that each worker handles more keys, which makes the busiest workers even more of a bottleneck

Question 15¶

Scenario: A company needs to transform JSON files in Cloud Storage, apply schema validation, and load into BigQuery. They want a visual, low-code solution that non-engineers can use. What should they use?

A. Write custom Python code with Apache Beam and run it on Dataflow, which works but requires engineers to build and maintain it B. Cloud Data Fusion with visual pipeline designer C. BigQuery JSON functions in SQL queries, which parse JSON but give non-engineers no visual interface for building a pipeline D. Spreadsheet import of the JSON files after converting them to rows, which has no schema validation and cannot scale to Cloud Storage

Question 16¶

Scenario: A company processes 100TB of data daily in batch jobs. Jobs run for 6 hours using Dataproc. They want to reduce processing time without changing code significantly. What should they do?

A. Process less data by sampling each day's input B. Run jobs less frequently, such as every other day, which gives each run more data and so makes each run longer C. Use smaller clusters to lower cost per hour, which reduces parallelism and makes a 100TB job take longer D. Optimize cluster configuration with SSDs, autoscaling, and enhanced flexibility mode

Question 17¶

Scenario: A data pipeline needs to process data that arrives with timestamps out of order. Some events arrive up to 1 hour late. The pipeline should include late data in the correct time windows. What should they configure in Dataflow?

A. Reject late data at the source and send it back to the partner for resubmission, which loses events that arrive within the hour B. Process all data as it arrives without windowing, which ignores event time and mixes late events into the wrong results C. Configure allowed lateness and watermark triggers in windowing strategy D. Ignore late data and let the next day's batch job correct the totals, which leaves the streaming results wrong in the meantime

Question 18¶

Scenario: A company wants to enrich streaming data with reference data. The reference data is 500 MB and updates daily. Enrichment should not slow down the streaming pipeline. What approach should they use?

A. Side inputs in Dataflow with periodically refreshed reference data, loaded once per worker instead of queried per event B. Join streams in memory by buffering the two inputs in worker state C. Enrich in a separate batch job that rewrites the output each day, which leaves the streaming data without enrichment until it runs D. Query BigQuery for each record as it arrives, which adds latency and cost per event

Question 19¶

Scenario: A company needs to process messages from Pub/Sub but must ensure exactly-once processing semantics. Some downstream systems are not idempotent. What should they implement?

A. Don't use Pub/Sub and read events directly from a database table, which offers no delivery guarantee B. Manual tracking in a spreadsheet of processed message IDs, which cannot keep pace C. Dataflow with checkpointing and idempotent sinks, or deduplication logic keyed on the Pub/Sub message ID before writing to the downstream system D. Accept duplicate processing and fix the totals in a nightly job

Domain 3: Store the Data (Questions 20-27)¶

Question 20¶

Scenario: A company needs to store 10PB of log data for 7 years. Data is rarely accessed but must be retrievable within a few hours when needed. They want to minimize storage cost. What should they use?

A. Cloud Storage Archive class with lifecycle policies and a bucket retention policy for the seven-year hold B. Bigtable with SSD nodes and a column family per month C. BigQuery with standard storage and a table per year D. Persistent Disk snapshots attached to a VM, which are meant for VM disks and cost far more than object storage at 10PB

Question 21¶

Scenario: A data warehouse in BigQuery has a large table (500TB) that's frequently queried by date range. Most queries filter on the last 30 days. Full table scans are expensive. How should they optimize?

A. Query the full table each time, relying on BigQuery's performance, which scans all 500TB and is charged for every byte read B. Use views instead of tables, which hide the filter logic but still scan the full table underneath on each query C. Partition by ingestion time or date field, cluster by frequently filtered columns D. Create many small tables, one for each month

Question 22¶

Scenario: A company needs to store user session data with TTL. Sessions expire after 30 minutes of inactivity. The data store should handle millions of reads/writes per second. What should they use?

A. BigQuery, which is an analytics warehouse with latency in the range of seconds and no per-key expiration of session rows B. Cloud SQL with a cleanup job that deletes old rows, which cannot scale to millions of reads and writes per second C. Cloud Memorystore (Redis) with key expiration D. Cloud Storage with object lifecycle rules, which expire objects by age in days and are not a low-latency key-value store

Question 23¶

Scenario: A company needs to store document data (JSON) with flexible schema. Documents vary in structure, and they need to query by various fields. Queries should return in milliseconds. What should they use?

A. Cloud Storage with a JSON object per document, which offers no field-level queries B. Cloud SQL with a JSON column, which works best with a defined schema C. BigQuery with nested and repeated fields, which suits analytics on JSON but has query latency in seconds, not milliseconds D. Firestore in Native mode, a document database with flexible schema and automatic field indexes

Question 24¶

Scenario: A company stores sensitive customer data in BigQuery. They need to ensure the data is encrypted with keys they control and that keys can be rotated without re-encrypting all data. What should they configure?

A. No encryption in the application, trusting the network perimeter, which does not meet the requirement for customer-controlled keys B. Default encryption with Google-managed keys, which protects data at rest but gives the customer no control over key rotation C. Client-side encryption before loading, in which the customer holds keys but BigQuery cannot query the encrypted columns D. Customer-managed encryption keys (CMEK) with automatic key rotation in Cloud KMS

Question 25¶

Scenario: A company needs to build a graph database to model social connections between users. They need to traverse relationships efficiently (friend-of-friend queries). What should they use?

A. Cloud Storage with edge lists stored as files, which holds the data but offers no traversal engine for friend-of-friend queries B. Firestore with a document per user listing friend IDs, which looks up one hop quickly but is not optimized for deep graph traversals C. Cloud SQL with join tables, which needs a self join for every hop D. BigQuery with recursive queries for simple graphs, or JanusGraph on Bigtable for large graph traversals

Question 26¶

Scenario: A data team needs to create a central catalog of all datasets across BigQuery, Cloud Storage, and Pub/Sub. They want to tag data with classifications (PII, confidential) and enable data discovery. What should they use?

A. Spreadsheet tracking of datasets by owner and sensitivity, which cannot scale or discover new data B. README files kept in each repository and bucket, which give no central search or tags C. Dataplex Universal Catalog (now called Knowledge Catalog), cataloging BigQuery, Cloud Storage and Pub/Sub, with aspects that mark data as PII or confidential D. Custom database for metadata with a web form for entries, which the team must build and host

Question 27¶

Scenario: A company needs to ensure referential integrity between tables in their data warehouse. Parent records should not be deleted if child records exist. They use BigQuery. What should they do?

A. BigQuery doesn't support primary or foreign key constraints at all B. Use BigQuery primary and foreign key constraints (for query optimization, not enforcement) or implement validation in ETL C. Don't validate relationships and rely on analysts to spot orphan records in reports D. Use Cloud SQL instead, which enforces foreign keys and blocks deletes, but changes the architecture and gives up warehouse-scale analytics

Domain 4: Prepare and Use Data for Analysis (Questions 28-33)¶

Question 28¶

Scenario: Data analysts need to create reports from BigQuery data. They prefer a drag-and-drop interface without writing SQL. Reports should refresh automatically. What tool should they use?

A. Looker Studio (formerly Data Studio) connected to BigQuery, with drag-and-drop report building and scheduled data refresh B. Export to Excel on a schedule and build the reports in spreadsheets, which needs manual refresh C. Command-line queries with the bq tool, which return results as text D. Write custom SQL queries and save them as views for each report

Question 29¶

Scenario: A data team needs to schedule and orchestrate complex data pipelines with dependencies between tasks. If one task fails, dependent tasks should not run. They need monitoring and alerting. What should they use?

A. Cloud Composer (Airflow) with DAGs defining task dependencies B. Cron jobs on a VM calling shell scripts in sequence, which run on a timer but have no dependency graph, retries, or alerting C. Cloud Scheduler alone triggering an HTTP endpoint for each task, which starts jobs on a schedule but does not track dependencies D. Manual execution by an engineer who checks each step before starting the next, which does not give monitoring or alerting

Question 30¶

Scenario: A company's data contains PII that needs to be masked before analysts can access it. Names and emails should be pseudonymized consistently (same input produces same output). What should they use?

A. Manual masking by an engineer who edits exported tables B. Encrypt entire tables with a key, which protects the data but leaves analysts unable to read any value for analysis C. Cloud DLP with deterministic tokenization (a keyed transformation) D. Delete PII columns from the tables, which removes the risk but also removes names and emails that analysis needs

Question 31¶

Scenario: A machine learning team needs to share features between multiple ML models. Features should be computed once and served consistently for both training and inference. What should they use?

A. Compute features in each model with its own preprocessing code, which duplicates work and causes training-serving skew B. Agent Platform Feature Store for centralized feature storage and serving C. A shared spreadsheet of feature definitions D. Database tables in Cloud SQL for each feature set, which store values but add no versioning or online and offline consistency

Question 32¶

Scenario: Business users need to explore data and create ad-hoc queries without knowing SQL. They should be able to ask questions in natural language. What should be enabled?

A. Gemini in BigQuery for natural language questions, with Looker's semantic layer and BI Engine accelerating interactive queries B. SQL training for all users, which takes time and does not allow plain-language questions C. Export to spreadsheets for pivot tables, which allows only small extracts D. Pre-built reports only, which limits users to questions someone already anticipated

Question 33¶

Scenario: A company needs to track data lineage - understanding where data comes from, how it's transformed, and where it's used. This is required for regulatory compliance. What should they implement?

A. Code comments in the pipeline source, which give no map of data flow B. Ignore lineage and rely on the data owners to answer questions, which does not meet a regulator's request for traceability C. Dataplex data lineage, with Dataflow publishing lineage to the Data Lineage API, from source to destination D. Manual documentation in a wiki with diagrams, which is out of date once a pipeline changes

Domain 5: Maintain and Automate Data Workloads (Questions 34-40)¶

Question 34¶

Scenario: A Dataflow streaming job needs to run 24/7. The team needs to be alerted if the job fails or if processing lag exceeds 5 minutes. What should they configure?

A. Cloud Monitoring alert policies on Dataflow job state and system lag (threshold 5 minutes), notifying the team by email, SMS, or PagerDuty B. No monitoring, since Dataflow restarts failed workers automatically C. Log analysis only, with a weekly search of the job logs for errors D. Check manually in the console at the start of each day

Question 35¶

Scenario: A BigQuery scheduled query fails occasionally due to transient errors. The team wants automatic retry before alerting. What should they configure?

A. No retries so that the first failure alerts the team, which sends alerts for transient errors that would succeed on a second run B. BigQuery scheduled queries with retry configuration and Cloud Monitoring alerts for persistent failures C. Don't use scheduled queries and run the SQL from a VM cron job, which gives no managed retry configuration or integrated alerting D. Manual re-runs by whoever sees the failure first, which are labor-intensive and slow to respond

Question 36¶

Scenario: A company wants to implement CI/CD for their Dataflow pipelines. Pipeline code should be tested before deployment, and deployments should be automated on merge to main. What should they implement?

A. Direct uploads of the pipeline code from a developer's machine B. No testing, with changes pushed straight to production when a reviewer approves, which risks breaking the running pipeline C. Manual deployments by an engineer who follows a runbook after each merge, which are slow and prone to error D. Cloud Build with unit tests, integration tests, and Dataflow Flex Template deployment

Question 37¶

Scenario: A BigQuery dataset contains tables that grow continuously. Old data (> 90 days) is rarely queried but must stay queryable in BigQuery, and its storage cost should fall automatically. What should they do?

A. Partition the table by date and leave old partitions unmodified, so BigQuery bills each at the lower long-term storage rate after 90 days B. Manual archival by an engineer who exports old tables each quarter and deletes them C. Rewrite every old partition into a new table each month, which resets the 90-day clock and keeps the active storage rate D. Delete old data after 90 days, which cuts cost but loses data that may be needed later

Question 38¶

Scenario: A Dataflow pipeline occasionally needs to be updated with new code. Updates should not lose in-flight data and should minimize processing interruption. What approach should they use?

A. Stop and restart the pipeline with the new code, which can lose or reprocess messages B. Dataflow in-place update for compatible changes, or drain and restart with a compatible schema for incompatible ones C. Delete and recreate the job from scratch, which loses in-flight data D. Run two pipelines in parallel on the same subscription, which splits messages between them and causes missed or duplicate processing

Question 39¶

Scenario: A data team wants to implement data quality checks. They need to validate that columns meet certain criteria (not null, within range, matching patterns) and alert when quality drops. What should they use?

A. Dataplex data quality rules with automated scans on a schedule and alerting when thresholds are breached B. Ad-hoc SQL checks that analysts run after loading when they suspect a problem C. Manual review of a sample of rows each week, which cannot alert when quality drops D. Ignore data quality and rely on downstream users to report issues, which lets bad data reach reports before it is caught

Question 40¶

Scenario: A company has multiple data pipelines managed by different teams. They want to implement access controls so teams can only modify their own pipelines but view others. What should they configure?

A. No access controls, so any team can fix another team's pipeline B. IAM roles bound at project or folder level, with custom roles for fine-grained access C. Shared credentials for all pipelines held in a team vault D. Everyone has admin access to the project so that no team is blocked, which lets any user change or delete another team's pipelines

Answer Key¶