Databricks Data Engineer Associate test
Twenty questions across the seven sections of the Databricks Data Engineer Associate exam, weighted roughly as the real exam is. Ingestion and transformation together carry over 40% of the score.
Databricks Intelligence Platform
Question 1. Which statement about Delta Lake is correct?
- A. It requires a dedicated database server
- B. It cannot support concurrent readers and writers
- C. It adds ACID transactions, schema enforcement and time travel over object storage
- D. It stores data only in memory
Show answer
Answer: C
Delta adds ACID transactions, schema enforcement and time travel on top of files in object storage. It does not replace object storage or require a separate database server.
Data Ingestion and Loading
Question 2. New JSON files land in cloud storage continuously. The pipeline must process each file exactly once without listing the whole directory each run. Which approach fits?
- A. Reload the entire directory on each run
- B. Incremental streaming ingestion with checkpointing
- C. Manually move files after processing
- D. Query the files directly without loading
Show answer
Answer: B
Incremental file ingestion with checkpointing tracks processed files and scales without repeated full directory listings. A full reload reprocesses everything.
Question 3. An ingestion job must not fail when a new optional column appears in source files. What should be configured?
- A. Schema evolution on the target table
- B. A hard-coded schema that rejects unknown columns
- C. Dropping the target table on each run
- D. Converting all columns to strings
Show answer
Answer: A
Schema evolution allows new columns to be added without breaking the write. Rejecting the files or hard-coding the schema causes failures on every source change.
Question 4. Raw source data must be preserved exactly as received for audit, while cleansing happens downstream. Which layer holds it?
- A. The gold layer
- B. The silver layer
- C. A temporary view
- D. The bronze layer
Show answer
Answer: D
The bronze layer holds raw ingested data as received. Silver holds cleansed data and gold holds business aggregates.
Question 5. A batch load re-runs after a partial failure and duplicates rows. What design prevents this?
- A. Append the data again and deduplicate in the dashboard
- B. Disable retries
- C. An idempotent merge keyed on a unique identifier
- D. Increase the cluster size
Show answer
Answer: C
An idempotent write keyed on a business or batch identifier, using merge, makes re-runs safe. Appending blindly duplicates on every retry.
Data Transformation and Modeling
Question 6. A silver table must apply the latest version of each record from a change feed, inserting new keys and updating existing ones. Which operation fits?
- A. MERGE INTO on the key
- B. INSERT INTO only
- C. TRUNCATE then INSERT
- D. A view over the change feed
Show answer
Answer: A
MERGE performs insert and update in one atomic operation keyed on the join condition. Separate insert and delete steps are neither atomic nor complete.
Question 7. Which is the correct reason to use a gold table rather than querying silver directly from dashboards?
- A. Gold tables use a different file format
- B. Gold tables pre-aggregate to business grain, cutting cost and enforcing consistent definitions
- C. Silver tables cannot be queried
- D. Gold tables do not require permissions
Show answer
Answer: B
Gold tables pre-aggregate to business-level grain, cutting query cost and enforcing consistent definitions across consumers.
Question 8. A transformation must keep history of changes to a dimension so past facts join to the attributes valid at the time. What is this called?
- A. Slowly changing dimension type 1
- B. A materialised view
- C. Slowly changing dimension type 2
- D. A temporary table
Show answer
Answer: C
A slowly changing dimension type 2 keeps historical versions with validity periods. Type 1 overwrites and loses history.
Question 9. A pipeline filters rows, then joins a large table, then aggregates. Which ordering generally performs best?
- A. Filter first, then join, then aggregate
- B. Join first, then filter, then aggregate
- C. Aggregate first, then join, then filter
- D. The order makes no difference in any engine
Show answer
Answer: A
Filtering before joining reduces the volume entering the join, which is normally the most expensive step. Joining first processes rows that are later discarded.
Question 10. Rows arriving with a null business key must be quarantined rather than loaded or silently dropped. What should the pipeline include?
- A. A filter that discards them
- B. A default key value
- C. No handling; they will be ignored
- D. A data quality rule routing failing rows to a quarantine table
Show answer
Answer: D
An explicit quality rule that routes failing rows to a quarantine table preserves them for investigation while keeping the target clean.
Working with Lakeflow Jobs
Question 11. A job has three tasks where the third must run only if both earlier tasks succeed. How should this be expressed?
- A. Three separate jobs scheduled minutes apart
- B. One job with task dependencies and success conditions
- C. A single notebook with all logic and no error handling
- D. A manual run after checking the first two
Show answer
Answer: B
Task dependencies in the job definition enforce ordering and success conditions. Separate scheduled jobs with guessed timings are fragile.
Question 12. A nightly job occasionally fails on a transient cloud storage error. What is the appropriate configuration?
- A. A bounded retry policy with alerting on final failure
- B. Unlimited retries with no alert
- C. Disabling the job on first failure permanently
- D. Running the job twice every night
Show answer
Answer: A
A bounded retry policy handles transient errors without masking persistent ones. Unlimited retries hide real failures and can run indefinitely.
Question 13. A job must start when an upstream job finishes rather than at a fixed time. What fits?
- A. A cron schedule 30 minutes later
- B. Manual start by an operator
- C. A trigger on the upstream job’s completion
- D. A continuous job that polls a table
Show answer
Answer: C
Triggering on completion of the upstream job removes timing guesswork. A fixed schedule with a safety margin wastes time and still races.
Governance and Security
Question 14. A team must query a table but must never see one sensitive column. What should be applied?
- A. A copy of the table without that column
- B. A column mask on the sensitive column
- C. A comment warning users not to select it
- D. Revoking access to the schema
Show answer
Answer: B
A column mask redacts the value for unprivileged principals while leaving the rest of the row usable. Copying the table without the column duplicates data and drifts.
Question 15. An auditor asks which tables a given job reads and writes. What provides this?
- A. The cluster event log
- B. The job’s cost report
- C. A manual read of the notebook code
- D. Table and job lineage in the governance catalog
Show answer
Answer: D
Lineage captured by the governance layer shows upstream and downstream relationships for tables and jobs. Reading notebook code by hand is unreliable and does not cover runtime behaviour.
Question 16. Access must be granted to a group rather than to individuals so leavers lose access automatically. Which principle does this follow?
- A. Group-based access management
- B. Row-level filtering
- C. Encryption at rest
- D. Schema evolution
Show answer
Answer: A
Group-based access management keeps entitlements tied to role membership, so identity lifecycle changes propagate automatically.
Implementing CI/CD
Question 17. Pipeline code must be promoted from development to production with review and traceability. What should be used?
- A. Copying notebooks between workspaces manually
- B. Exporting and emailing the notebooks
- C. Version control with review and an automated deployment pipeline
- D. Editing directly in production
Show answer
Answer: C
Version control with pull request review and an automated deployment pipeline gives both review and traceability. Manual copying leaves no record.
Question 18. The same pipeline must run against development and production data without editing the code. What should differ?
- A. A separate copy of the code per environment
- B. Parameterised configuration for catalog, schema and storage paths
- C. Hard-coded paths changed before each deployment
- D. A different programming language per environment
Show answer
Answer: B
Environment-specific configuration such as catalog, schema and paths should be parameterised, leaving the code identical across environments.
Troubleshooting, Monitoring, and Optimization
Question 19. A job that used to finish in 20 minutes now takes two hours, with no code change. What should you examine first?
- A. Input data volume and file layout, including small file counts and skew
- B. The workspace theme settings
- C. The notebook’s comment density
- D. The job’s display name
Show answer
Answer: A
Input data volume and file layout, such as an explosion of small files or skewed partitions, commonly explain gradual slowdowns without code change.
Question 20. A stage shows a few tasks running far longer than the rest. What is this symptom called?
- A. Schema drift
- B. Checkpoint corruption
- C. Broadcast failure
- D. Data skew
Show answer
Answer: D
Data skew concentrates disproportionate data in a few partitions, so a handful of tasks dominate runtime. Shuffling and caching are mechanisms, not the symptom.
How did you do?
Sixteen or more correct suggests you are close. Below fourteen, the section guides here will help most. Databricks does not publish a passing score for this exam, so treat quoted numbers elsewhere with caution.