Free DP-750 sample questions with answers
Here are five free DP-750 sample questions covering all four skill areas, with the answer and the reasoning behind it. Try each one before opening the answer. These are original scenario questions written against the official skills outline, not questions from the real exam.
Question 1. A transformation notebook runs once every night for about 40 minutes, with no interactive users. The team wants to pay for compute only while the run is active and not manage a long-lived cluster. Which compute choice fits best?
- A. An all-purpose cluster with auto-termination disabled
- B. Job compute, or serverless compute for jobs, that exists only for the run
- C. A SQL warehouse kept running overnight
- D. A shared compute resource that analysts use during the day
Show answer
Answer: B
Job compute (or serverless compute for jobs) is created for the run and released when it finishes, so cost follows the run. An all-purpose cluster left running accrues charges between runs. A SQL warehouse is designed for SQL queries and BI workloads, not a Python notebook pipeline. Shared compute for interactive users has the same idle-cost problem as all-purpose compute.
Want more questions like this? Full DP-750 practice tests →
Question 2. A table holds sales for all regions. Members of the group emea-analysts must see only rows where region is EMEA, while other groups see everything. You do not want to maintain separate copies of the table. What should you implement in Unity Catalog?
- A. A separate table per region with its own grants
- B. A column mask on the region column
- C. A row filter on the table based on group membership
- D. A GRANT SELECT to emea-analysts on the table
Show answer
Answer: C
A row filter attached to the table evaluates a function per user or group and returns only permitted rows, so one table serves everyone. Copying the table per region creates the duplication you wanted to avoid. A column mask changes column values and does not remove rows. Granting SELECT only controls whether a principal can read the table at all, not which rows.
Want more questions like this? Full DP-750 practice tests →
Question 3. New JSON files land continuously in a cloud storage folder, millions over time. The pipeline must pick up only new files, tolerate new columns appearing, and not rescan the whole folder on each run. Which approach should you use?
- A. A CREATE TABLE AS SELECT over the whole folder on every run
- B. A batch spark.read of the folder at each run, then append
- C. A one-time load into a Delta table
- D. Auto Loader reading the folder as a stream, with schema evolution enabled
Show answer
Answer: D
Auto Loader tracks which files have been processed, scales file discovery, and supports schema inference and evolution, so each run handles only new files. Re-reading the folder with a batch read or CTAS each run reprocesses everything. Listing files by hand in the notebook does not scale and duplicates what Auto Loader does. A one-time load does not handle files that keep arriving.
Want more questions like this? Full DP-750 practice tests →
Question 4. In a Lakeflow Spark Declarative Pipeline, rows where order_id is null must be removed from the target table, and the pipeline must keep running. How should you define the data quality rule?
- A. An expectation on order_id IS NOT NULL with the drop-row action
- B. An expectation on order_id IS NOT NULL with no violation action
- C. An expectation on order_id IS NOT NULL with the fail-update action
- D. A VACUUM command after each update
Show answer
Answer: A
An expectation with the drop-row action discards violating rows, records the counts in pipeline metrics and lets the update continue. An expectation with no action only records violations and keeps the bad rows. A fail-update action stops the update when a row violates the rule. VACUUM removes old data files from a Delta table and has no role in row validation.
Want more questions like this? Full DP-750 practice tests →
Question 5. A Lakeflow Job has five tasks. The first three succeeded, task four failed because the source system was briefly unavailable, and task five depends on task four. The source is now back. What is the most efficient way to finish the run?
- A. Run the whole job again from the first task
- B. Delete the job and recreate it from the bundle
- C. Repair the run so only the failed task and its dependants run again
- D. Change the schedule to run an hour later
Show answer
Answer: C
A repair run re-executes only the failed task and the tasks that depend on it, reusing the successful results of the earlier tasks. Running the whole job again repeats work already done. Deleting and recreating the job loses history and fixes nothing. Moving the schedule does not recover the current failed run.
Want more questions like this? Full DP-750 practice tests →
How did you do?
Questions 3 and 4 trip up engineers who have used notebooks but not Lakeflow Spark Declarative Pipelines: the exam names Auto Loader and expectations explicitly, and the details of what each violation action does matter. Question 2 is a good test of whether you know the difference between a grant, a row filter and a column mask. If any of these were slow, read the data processing guide or the Unity Catalog security guide.
Question 5 is about the operational side. The outline lists repair, restart, stop and run as separate job actions, and the exam expects you to choose the right one.
What these questions have in common
None of them asks for a definition. Each states a requirement and offers options that all sound plausible. One detail decides it: pay only for the run, one table for everyone, only new files, keep running, only the failed part. Underline that constraint before reading the options.
For a longer check, take the 20-question practice test.