Professional Data Engineer Sample

Questions
The Data Engineer sample questions will familiarize you with the format of exam questions
and example content that may be covered on the exam.

The sample questions do not represent the range of topics or level of difficulty of
questions presented on the exam. Performance on the sample questions should not be
used to predict your Data Engineer exam result.

Registration

First Name *

Francisco

Last Name *

Mussari

Primary Email *

fmussari@gmail.com

Recovery Email

fmussari@gmail.com
Organization (Employer or School) *

None

Organization email (an email associated with your current organization)

Country *

Venezuela

Primary Relationship to Google *

Customer

Send me offers, updates and useful tips for getting the most out of Google Cloud *
training and certification products and services.

No

Question 1
 You are working on optimizing BigQuery for a query that is run repeatedly on a
single table. The data queried is about 1 GB, and some rows are expected to
change about 10 times every hour. You have optimized the SQL statements as
much as possible. You want to further optimize the query's performance.
What should you do?

A. Create a materialized view based on the table, and query that view.

B. Enable caching of the queried data so that subsequent queries are faster.

C. Create a scheduled query, and run it a few minutes before the report has to be
created.

D. Reserve a larger number of slots in advance so that you have maximum compute
power to execute the query.

Correct answer

A. Create a materialized view based on the table, and query that view.

Feedback

A: Option A is correct because materialized views periodically cache the results of a query
for increased performance. Materialized views are suited to small datasets that are
frequently queried. When underlying table data changes, the materialized view invalidates
the affected portions and re-reads them.

B: Option B is not correct because caching is automatically enabled but is not performant
when the underlying data changes.

C: Option C is not correct because scheduled queries let you schedule recurring queries
but do not provide specific performance
optimizations. Also, running a query too early could use old/stale data.

D: Option D is not correct because reserving more slots guarantees the availability of
BigQuery slots but does not improve performance.

https://cloud.google.com/… https://cloud.google.com/…

https://cloud.google.com/…

Question 2
 Several years ago, you built a machine learning model for an ecommerce
company. Your model made good predictions. Then a global pandemic
occurred, lockdowns were imposed, and many people started working from
home. Now the quality of your model has degraded. You want to improve the
quality of your model and prevent future performance degradation. What
should you do?

A. Retrain the model with data from the first 30 days of the lockdown.

B. Monitor data until usage patterns normalize, and then retrain the model.

C. Retrain the model with data from the last 30 days. After one year, return to the
older model.

D. Retrain the model with data from the last 30 days. Add a step to continuously
monitor model input data for changes, and retrain the model.

Correct answer

D. Retrain the model with data from the last 30 days. Add a step to continuously
monitor model input data for changes, and retrain the model.

Feedback

A: Option A is not correct because retraining based on the data from the first 30 days of
the lockdown might only be useful for predictions during similar lockdowns and not for
regular periods.
B: Option B is not correct because usage patterns might have changed permanently and
might continue to change in the future.
C: Option C is not correct because the older model might not be indicative of user behavior
after a year.
D: Option D is correct because the data used to build the original model is no longer
relevant. Retraining the model with recent data from the last 30 days will improve the
predictions. To keep a watch on future data drifts, monitor the incoming data.

https://cloud.google.com/…

Question 3
 A new member of your development team works remotely. The developer will
write code locally on their laptop, which will connect to a MySQL instance on
Cloud SQL. The instance has an external (public) IP address. You want to
follow Google-recommended practices when you give access to Cloud SQL to
the new team member. What should you do?

A. Ask the developer for their laptop's IP address, and add it to the authorized
networks list.

B. Remove the external IP address, and replace it with an internal IP address. Add
only the IP address for the remote developer's laptop to the authorized list.

C. Give instance access permissions in Identity and Access Management (IAM),

and have the developer run Cloud SQL Auth proxy to connect to a MySQL instance.

D. Give instance access permissions in Identity and Access Management (IAM),

change the access to "private service access" for security, and allow the
developer to access Cloud SQL from their laptop.

Correct answer

C. Give instance access permissions in Identity and Access Management (IAM), and
have the developer run Cloud SQL Auth proxy to connect to a MySQL instance.

Feedback

A: Option A is not correct, because although adding an authorized networks list is

possible, it is more effort to track it and also less secure.
B: Option B is not correct, because if you remove the external IP address, access for those
who work remotely will be more complicated because they also have to be within private
RFC 1918 address space.
C: Option C is correct because the recommended approach is to use Cloud SQL Auth
proxy. Permissions can be controlled by IAM. You don't need to track authorization lists for
changing user IP addresses.
D: Option D is not correct because private service access will require access from a
private RFC 1918 address space, which might not be available to developers who work
remotely.

https://cloud.google.com/… https://cloud.google.com/…

https://cloud.google.com/… https://cloud.google.com/…

https://codelabs.develope…

Question 4
 Your Cloud Spanner database stores customer address information that is
frequently accessed by the marketing team. When a customer enters the
country and the state where they live, this information is stored in different
tables connected by a foreign key. The current architecture has performance
issues. You want to follow Google-recommended practices to improve
performance. What should you do?

A. Create interleaved tables, and store states under the countries.

B. Denormalize the data, and have a row for each state with its corresponding
country.

C. Retain the existing architecture, but use short, two-letter codes for the countries
and states.

D. Combine the countries in a single cell's text, for example "country:state1,state2,

…" and when required, split the data.

Feedback

A: Option A is correct because Cloud Spanner supports interleaving that guarantees data
being stored in the same split, which is performant when you need a strong data locality
relationship.
B: Option B is not correct because denormalizing is not a preferred approach in relational
databases. It leads to multiple rows with repeated data.
C: Option C is not correct because reducing the size of the fields to short names will have
lower impact because the data access and joins will be a bigger performance issue.
D: Option D is not correct because packing multiple types of data into the same cell is not
recommended for relational databases.

https://cloud.google.com/…

Question 5
 Your company runs its business-critical system on PostgreSQL. The system is
accessed simultaneously from many locations around the world and supports
millions of customers. Your database administration team manages the
redundancy and scaling manually. You want to migrate the database to
Google Cloud. You need a solution that will provide global scale and
availability and require minimal maintenance. What should you do?

A. Migrate to BigQuery.

B. Migrate to Cloud Spanner.

C. Migrate to a Cloud SQL for PostgreSQL instance.

D. Migrate to bare metal machines with PostgreSQL installed.

Correct answer

B. Migrate to Cloud Spanner.

Feedback

A: Option A is not correct because BigQuery doesn’t support global scale. BigQuery also
isn’t the best option for migrating a transactional database like PostgreSQL because it is
more analytics-focused.
B: Option B is correct because Cloud Spanner provides a global-scale, highly available
database that supports relational data.
C: Option C is not correct because Cloud SQL options are regional and have less scalability
compared to Cloud Spanner.
D: Option D is not correct because running PostgreSQL on bare metal machines requires a
greater maintenance effort.

https://cloud.google.com/…

Question 6
 Your company collects data about customers to regularly check their health
vitals. You have millions of customers around the world. Data is ingested at
an average rate of two events per 10 seconds per user. You need to be able to
visualize data in Bigtable on a per user basis. You need to construct the
Bigtable key so that the operations are performant. What should you do?

A. Construct the key as user-id#device-id#activity-id#timestamp.

B. Construct the key as timestamp#user-id#device-id#activity-id.

C. Construct the key as timestamp#device-id#activity-id#user-id.

D. Construct the key as user-id#timestamp#device-id#activity-id.

Correct answer

A. Construct the key as user-id#device-id#activity-id#timestamp.

Feedback

A: Option A is correct because the design does not monotonically increase, thus avoiding
hotspots.
B: Option B is not correct because it monotonically increases, thus causing hotspots.
C: Option C is not correct because it monotonically increases, thus causing hotspots.
D: Option D is not correct because it monotonically increases, thus causing hotspots.

https://cloud.google.com/…

Question 7
 Your company is hiring several business analysts who are new to BigQuery.
The analysts will use BigQuery to analyze large quantities of data. You need to
control costs in BigQuery and ensure that there is no budget overrun while you
maintain the quality of query results. What should you do?

A. Set a customized project-level or user-level daily quota to acceptable values.

B. Reduce the data in the BigQuery table so that the analysts query less data, and
then archive the remaining data.

C. Train the analysts to use the query validator or --dry_run to estimate costs so that
the analysts can self-regulate usage.

D. Export the BigQuery daily costs, and visualize the data on Looker on a per-
analyst basis so that the analysts can self-regulate usage.

Feedback

A: Option A is correct because if you have multiple BigQuery projects and users, you can
manage costs by requesting a custom quota that specifies a limit on the amount of query
data processed per day.
B: Option B is not correct because giving only partial data to the analysts might not
produce accurate query results.
C: Option C is not correct because costs could still overrun budgets. This approach
assumes that analysts always follow guidelines.
D: Option D is not correct because your costs could still overrun budgets. This approach
also assumes that the analysts look at the charts daily and adjust their behavior.

https://cloud.google.com/…

Question 8
 Your Bigtable database was recently deployed into production. The scale of
data ingested and analyzed has increased significantly, but the performance
has degraded. You want to identify the performance issue. What should you
do?

A. Use Key Visualizer to analyze performance.

B. Use Cloud Trace to identify the performance issue.

C. Add logging statements into the code to see which inserts cause the delay.

D. Add more nodes to the cluster to see if that resolves the performance issue.

Correct answer

A. Use Key Visualizer to analyze performance.

Feedback

A: Option A is correct because Key Visualizer for Bigtable generates visual reports for your
tables that detail your usage based on the row keys that you access, show you how
Bigtable operates, and can help you troubleshoot performance issues.
B: Option B is not correct because Cloud Trace is used to debug latency in applications.
C: Option C is not correct because adding logging statements won't help you understand
the performance issues within Bigtable.
D: Option D is not correct because adding more nodes might improve the performance, but
the database could continue to have performance issues if the keys are not designed well.

https://cloud.google.com/…

Question 9
 Your company is moving your data analytics to BigQuery. Your other
operations will remain on-premises. You need to transfer 800 TB of historic
data. You also need to plan for 30 Gbps of daily data transfers that must be
appended for analysis the next day. You want to follow Google-recommended
practices to transfer your data. What should you do?

A. As early as possible every day, use Cloud VPN to transfer the existing data over
the internet.

B. Use a Transfer Appliance to move the existing data to Google Cloud. Use
Cloud VPN to transfer data daily.

C. Use a Transfer Appliance to move the existing data to Google Cloud.. Use VPC
Network Peering to transfer data daily.

D. Use a Transfer Appliance to move the existing data to Google Cloud. Set up a
Dedicated or Partner Interconnect for daily transfers.

Correct answer

D. Use a Transfer Appliance to move the existing data to Google Cloud. Set up a
Dedicated or Partner Interconnect for daily transfers.

Feedback

A: Option A is not correct because the internet in general will have less stability and much
lower speed. Transferring large amounts of data is not viable.
B: Option B is not correct because Cloud VPN is useful for data transfers at a rate of a few
Gbps (1.5 Gbps to 3 Gbps).
C: Option C is not correct because VPC Network Peering is used for data transfers within
Google Cloud Organizations.
D: Option D is correct because using a Transfer Appliance is recommended to transfer
hundreds of terabytes of data. For large data transfers that occur regularly, a dedicated,
hybrid networking connection is recommended.

https://cloud.google.com/… https://cloud.google.com/…

https://cloud.google.com/…

Question 10
 Your team runs Dataproc workloads where the worker node takes about 45
minutes to process. You have been exploring various options to optimize the
system for cost, including shutting down worker nodes aggressively. However,
in your metrics you see that the entire job takes even longer. You want to
optimize the system for cost without increasing job completion time. What
should you do?

A. Set a graceful decommissioning timeout greater than 45 minutes.

B. Rewrite the processing in Cloud Data Fusion, and run the job automatically.

C. Rewrite the processing in Dataflow, and use stream processing of the same
data.

D. Increase the number of vCPUs on each worker node so that the processing
finishes sooner.

Correct answer

A. Set a graceful decommissioning timeout greater than 45 minutes.

Feedback

A: Option A is correct because graceful decommissioning will finish work in progress on a

worker node before it is removed from the Dataproc cluster.
B: Option B is not correct because rebuilding the data pipeline in Cloud Data Fusion will
increase effort, cost, and time.
C: Option C is not correct because rewriting the code in Dataflow will increase effort, cost,
and time.
D: Option D is not correct because increasing the number of vCPUs will greatly increase
the cost.

https://cloud.google.com/… https://cloud.google.com/…

Question 11
 Your customer has a SQL Server database that contains about 5 TB of data in
another public cloud. You expect the data to grow to a maximum of 25 TB.
The database is the backend of an internal reporting application that is used
once a week. You want to migrate the application to Google Cloud to reduce
administrative effort while keeping costs the same or reducing them. What
should you do?

A. Migrate the database to Bigtable.

B. Migrate the database to Cloud Spanner.

C. Install SQL Server on a Compute Engine VM.

D. Migrate the database to SQL Server in Cloud SQL.

Correct answer

D. Migrate the database to SQL Server in Cloud SQL.

Feedback

A: Option A is not correct because Bigtable is a NoSQL database and is not a suitable
destination from a SQL Server source.
B: Option B is not correct because Cloud Spanner will be costlier than Cloud SQL. Although
Spanner has global availability, it is unnecessary for this application requirement.
C: Option C is not correct because installing a custom SQL Server instance on a Compute
Engine VM will require more administrative effort.
D: Option D is correct because Cloud SQL provides managed MySQL, PostgreSQL, and SQL
Server databases, which will reduce administrative effort. Twenty-five TB can be
accommodated efficiently on Cloud SQL.

https://cloud.google.com/… https://cloud.google.com/…

https://cloud.google.com/… https://cloud.google.com/…

Question 12
 Your IT team uses BigQuery for storing structured data. Your finance team
recently moved to Google Workspace Enterprise edition from a standalone,
desktop-based spreadsheet processor. When the finance team needs data
insights, the IT team runs a query on BigQuery, exports the data to a CSV file,
and sends the file as an email attachment to the finance team members. You
want to improve the process while you retain familiar methods of data
analysis for the finance team. What should you do?

A. Run the query in BigQuery, and give the finance team access to the results view,
which can be analyzed.

B. Run the query in BigQuery, and give the finance team access to the data
visualizations in Google Data Studio.

C. Run the query in BigQuery, export the data to CSV, upload the file to a Cloud
Storage bucket, and share the file with the finance team.

D. Run the query in BigQuery, and save the results to a Google Sheets shared
spreadsheet that can be accessed and analyzed by the finance team.

Feedback

A: Option A is not correct because the finance team will have to be given Google Cloud
access and be trained on using BigQuery, which is not a familiar method.
B: Option B is not correct because only giving the visualizations on Data Studio won't let
the finance teams analyze the data.
C: Option C is not correct because the finance team will have to be given Google Cloud
access and be trained on using Cloud Storage, which is not a familiar method.
D: Option D is correct because Connected Sheets gives you a direct and easy way to share
BigQuery data through Google Sheets.

https://cloud.google.com/… https://cloud.google.com/…

https://www.youtube.com…

Question 13
 Your scooter-sharing company collects information about their scooters, such
as location, battery level, and speed. The company visualizes this data in real
time. To guard against intermittent connectivity, each scooter sends repeats
of certain messages within a short interval. Occasional data errors have been
noticed. The messages are received in Pub/Sub and stored in BigQuery. You
need to ensure that the data does not contain duplicates and that erroneous
data with empty fields is rejected. What should you do?

A. Store the data in BigQuery, and run delete queries on erroneous and duplicate
data.

B. Use Dataflow to subscribe to Pub/Sub, process the data, and store the data
in BigQuery.

C. Use Kubernetes to create a microservices application that can remove

duplicates and erroneous data. Then insert the data into BigQuery.

D. Create an application on Compute Engine with Managed Instance Groups that

can remove duplicates and erroneous data. Then insert the data into BigQuery.

Feedback

A: Option A is not correct because directly storing data in BigQuery could cause data to be
overwritten, and erroneous data could be present before it is deleted. Workarounds within
BigQuery to circumvent these concerns would cost more effort, time, and money.
B: Option B is correct because Dataflow is the recommended data processing product for
streaming data. Dataflow can be programmed to remove duplicates, delete empty fields,
and perform other custom data processing.
C: Option C is not correct because creating a custom application for streaming processing
on Kubernetes is a significant effort and is not recommended.
D: Option D is not correct because creating a custom application for streaming processing
on Compute Engine is a significant effort and is not recommended.

https://cloud.google.com/…

Question 14
 Your cryptocurrency trading company visualizes prices to help your customers
make trading decisions. Because different trades happen in real time, the
price data is fed to a data pipeline that uses Dataflow for processing. You
want to compute moving averages. What should you do?

A. Use hopping windows in Dataflow.

B. Use session windows in Dataflow.

C. Use tumbling windows in Dataflow.

D. Use Dataflow SQL, and compute averages grouped by time.

Feedback

A: Option A is correct because you use hopping windows to compute moving averages.
B: Option B is not correct because session windows are not used to calculate moving
averages.
C: Option C is not correct because tumbling windows are not used to calculate moving
averages.
D: Option D is not correct because grouping by time alone does not give you a moving
average.

https://cloud.google.com/…

Question 15
 You are building the trading platform for a stock exchange with millions of
traders. Trading data is written rapidly. You need to retrieve data quickly to
show visualizations to the traders, such as the changing price of a particular
stock over time. You need to choose a storage solution in Google Cloud. What
should you do?

A. Use Bigtable.

B. Use Firestore.

C. Use Cloud SQL.

D. Use Memorystore.

Feedback

A: Option A is correct because Bigtable is the recommended database for time series data
that requires high throughput reads and writes.
B: Option B is not correct because Firestore does not have the high throughput capabilities
that are suitable for time-series data.
C: Option C is not correct because Cloud SQL does not the have high throughput
capabilities that are suitable for time-series data.
D: Option D is not correct because Memorystore is a fast in-memory database that is not
suitable for persistently storing large amounts of data.

https://cloud.google.com/…

Question 16
 Your customer uses Hadoop and Spark to run data analytics on-premises. The
main data is stored in hard disks that are centrally accessed. Your customer
needs to migrate their workloads to Google Cloud efficiently while considering
scalability. You want to select an architecture that requires minimal effort.
What should you do?

A. Use Dataproc to run Hadoop and Spark jobs. Move the data to Cloud Storage.

B. Use Dataflow to recreate the jobs in a serverless approach. Move the data to
Cloud Storage.

C. Use Dataproc to run Hadoop and Spark jobs. Retain the data on a Compute
Engine VM with an attached persistent disk.

D. Use Dataflow to recreate the jobs in a serverless approach. Retain the data on a
Compute Engine VM with an attached persistent disk.

Feedback

A: Option A is correct because Dataproc is a fully managed service for hosting open
source distributed processing platforms, such as Apache Spark, Presto, Apache Flink and
Apache Hadoop on Google Cloud. Cloud Storage is the preferred storage option for all
persistent storage needs.
B: Option B is not correct because using Dataflow requires you to rewrite all the jobs.
C: Option C is not correct because storing the centrally accessed data on persistent disks
is not recommended.
D: Option D is not correct because using Dataflow requires you to rewrite all the jobs.
Storing the centrally accessed data on persistent disks is not recommended.

https://cloud.google.com/… https://cloud.google.com/…

https://cloud.google.com/…

Question 17
 You are on a team of analysts who work with BigQuery and are already
proficient in SQL. Your team needs to build a multi-label machine learning
classification model that uses data in BigQuery. There are 6,000 rows of data
in your training dataset. The inferences could be one of 200 possible labels.
You want to create a high accuracy model. What should you do?

A. Use BigQuery ML to create a model.

B. Export the data to CSV files. Use TensorFlow to build a model.

C. Connect the data from BigQuery to AutoML, and build the model in AutoML.

D. Connect to the data through AI notebooks, and build your model interactively.

Correct answer

C. Connect the data from BigQuery to AutoML, and build the model in AutoML.

Feedback

A: Option A is not correct because BigQuery ML requires a lot of data to build an accurate
model, and you don't have much data.
B: Option B is not correct because there isn't enough data to build a custom model in
TensorFlow.
C: Option C is correct because the amount of data is relatively low and also varied. A
model built using only this data wouldn't be accurate. AutoML is appropriate because it
uses transfer learning based on other similar data.
D: Option D is not correct because there isn't enough data to build a custom model with AI
Notebooks.

https://cloud.google.com/…

Question 18
 You used a small amount of data to build a machine learning model that gives
you good inferences during testing. However, the results show more errors
when real-world data is used to run the model. No additional data can be
collected for testing. You want to get a more accurate view of the model's
capability. What should you do?

A. Reduce the amount of data to improve the model.

B. Cross-validate the data, and re-run the model building process.

C. Create feature crosses that will add new columns to increase the data.

D. Duplicate the data twice to increase the data, and re-run the model building
process.

Feedback

A: Option A is not correct because this model is not underfitting.

B: Option B is correct because this model appears to be overfitting. Using cross-validation
will run the validation on multiple folds of the data, which reduces the overfitting.
C: Option C is not correct because adding new columns will not reduce the overfitting.
D: Option D is not correct because duplicating the data will not reduce the overfitting.

https://developers.google.…

Question 19
 Your organization has been collecting information for many years about your
customers, including their address and credit card details. You plan to use this
customer data to build machine learning models on Google Cloud. You are
concerned about private data leaking into the machine learning model. Your
management is also concerned that direct leaks of personal data could
damage the company's reputation. You need to address these concerns about
data security. What should you do?

A. Remove all the tables that contain sensitive data.

B. Use libraries like SciPy to build the ML models on your local computer.

C. Remove the sensitive data by using the Cloud Data Loss Prevention (DLP)
API.

D. Identify the rows that contain sensitive data, and apply SQL queries to remove
only those rows.

Feedback

A: Option A is not correct because removing data, such as entire tables, could reduce the
effectiveness of the resulting model.
B: Option B is not correct because building machine learning models on individual
computers is not a viable approach when it involves large amounts of data.
C: Option C is correct because Cloud DLP is the recommended approach to redact, mask,
tokenize, and transform text and images to help protect data privacy.
D: Option D is not correct because removing data, such as full rows, could reduce the
effectiveness of the resulting model.

https://cloud.google.com/…

Question 20
 Your healthcare application has a backend system that accepts event data
directly from IoT devices. Recent increases of the application's users and
devices are causing a sudden influx of data that overwhelms the system. You
need to redesign the data pipeline to ensure that all data is processed and
that no events are lost. You want to follow Google-recommended practices.
What should you do?

A. Use Kafka with pull mode.

B. Use Pub/Sub with pull mode.

C. Use Pub/Sub with push mode.

D. Run Cloud Scheduler at fixed intervals.

Feedback

A: Option A is not correct because Kafka is not a managed solution in Google Cloud. The
Google-recommended option is Pub/Sub, a fully managed, serverless solution.
B: Option B is correct because pull mode allows new event data to be pulled for
processing on demand when the previous data is processed. Pub/Sub will absorb and
retain new events in the interim without losing them.
C: Option C is not correct because Pub/Sub in push mode could continue to overwhelm
the system.
D: Option D is not correct because new event data should be pulled for processing when
the previous processing is completed, and that is not expected to be at fixed intervals.

https://cloud.google.com/… https://cloud.google.com/…

This form was created inside of Google.com. Privacy & Terms

Forms