Databricks Certified Data Engineer Professional practice questions

Question 1

A platform team has decided to put its Databricks jobs and pipelines under source control with code review and CI/CD, and an architect proposes using Databricks Asset Bundles (Declarative Automation Bundles). What do bundles primarily provide for this goal?

Accepted Answer

An end-to-end, source-file definition of a project that describes resources such as jobs and pipelines for test and deploy

Answer

A managed Git host inside the workspace that replaces GitHub and automatically runs the unit tests for every pull request

Answer

A runtime cache that speeds up notebook execution by pinning compiled artifacts, unrelated to deployment or source control

Answer

A one-time export of existing jobs to JSON that cannot be redeployed and is intended only for archival backup purposes only

Question 2

An engineer has decided to speed up a large Delta MERGE by relying on dynamic file pruning to skip non-matching files, but pruning is not firing on the non-Photon cluster. Per the Databricks dynamic file pruning documentation, what explains this?

Accepted Answer

Dynamic file pruning in MERGE, UPDATE, and DELETE requires Photon-enabled compute, which is missing here

Answer

Dynamic file pruning only works on partitioned tables, so a non-partitioned target table is ineligible

Answer

Dynamic file pruning needs the table Z-ordered first, and a MERGE cannot read an unordered file layout

Answer

Dynamic file pruning is disabled whenever a WHERE clause is present, as a MERGE supplies its own filter

Question 3

An MLOps team is cleansing an ARRAY of raw string tags inside a Spark SQL transformation and wants to apply per-element logic without boilerplate or exploding the array, so the engineer must explain how a higher-order function and its lambda divide the work. Per the Databricks higher-order functions documentation, which option is correct?

Accepted Answer

The higher-order function takes the array and dictates the result, delegating to a lambda how to process each item in the array.

Answer

The lambda takes the whole array and dictates the final result, while the higher-order function only handles one element at a time.

Answer

The higher-order function first explodes the array into rows, then a lambda recombines those rows back into a single array column.

Answer

Lambdas only operate on MAP columns, so the array of tags must be cast to a MAP before any per-element cleansing logic runs.

Question 4

A team is evaluating Databricks Asset Bundles to bring software-engineering practices to its data and AI project, and the lead engineer must summarize what a bundle fundamentally provides. Which description matches?

Accepted Answer

A bundle describes Databricks resources such as jobs and pipelines as source files.

Answer

A bundle is only a wrapper for notebook source code and cannot define jobs or pipelines.

Answer

A bundle is a runtime cluster policy that governs autoscaling limits for interactive compute.

Answer

A bundle is a managed Git provider that replaces an external version-control system entirely.

Question 5

An architect is tasked with enabling customer-managed keys so that notebooks, secrets, and SQL query data held in the Databricks control plane are encrypted under a key the company controls. Which customer-managed key use case covers this data?

Accepted Answer

The managed services use case, covering control plane data such as notebooks, secrets, and SQL query data.

Answer

The workspace storage use case, covering the workspace storage bucket (DBFS root) and the classic compute EBS volumes.

Answer

An external location storage credential that Unity Catalog rotates automatically on every individual user query.

Answer

AWS PrivateLink encryption applied to the control-plane API traffic between the client network and Databricks.

Question 6

A team is configuring job notifications for a Structured Streaming task and must explain the timing semantics of the streaming backlog notification so on-call is not flooded with messages. Per the job notifications documentation, which option matches?

Accepted Answer

A notification fires when the 10-minute average backlog exceeds the threshold, then Databricks waits 30 minutes.

Answer

A notification fires the instant any single micro-batch backlog crosses the threshold, with no averaging window ever applied.

Answer

A notification fires every 10 minutes on a fixed schedule regardless of whether the defined backlog threshold was crossed.

Answer

A notification fires once and then stays permanently suppressed for the entire remaining lifetime of the streaming query.

Question 7

A platform team is rolling out attribute-based access control in Unity Catalog so that protection follows governed tags as new tables are classified, rather than editing grants table by table. Which TWO statements about ABAC and governed tags are correct?

Accepted Answer

ABAC determines access by evaluating attributes expressed as governed tags in policy conditions, with policies attached at catalog, schema, or table level and evaluated dynamically.

Accepted Answer

Governed tags are account-level tags with enforced rules where you define the allowed keys and values and control which users and groups are able to assign them to objects.

Answer

ABAC requires the table owner to run ALTER TABLE ... SET ROW FILTER for each governed tag before the policy that references that tag is able to take any effect.

Answer

Governed tags can be assigned or modified by any account user, and that openness is precisely what makes the tagged data broadly discoverable across the whole organization.

Question 8

A platform team is configuring Auto Loader for a long-lived JSON feed where upstream producers occasionally add new fields and the lead engineer is reviewing the documented values for cloudFiles.schemaEvolutionMode. Per the Auto Loader documentation, which TWO options are valid values?

Accepted Answer

addNewColumns - documented as a supported value of cloudFiles.schemaEvolutionMode that automatically widens the inferred schema when previously-unseen columns appear in new files. The published Databricks data-engineering documentation lists this as the canonical, supported behavior.

Accepted Answer

rescue - documented as a supported value of cloudFiles.schemaEvolutionMode that captures unexpected columns into the rescued data column instead of aborting the micro-batch on schema drift. The published Databricks data-engineering documentation lists this as the canonical, supported behavior.

Answer

softReject - a documented schemaEvolutionMode value that quarantines rows that have new columns but otherwise applies the prior schema, available since the recent runtime cycle on Databricks LTS clusters.

Answer

auditOnly - a documented schemaEvolutionMode value that records each schema drift event into a Unity Catalog audit table without applying the new columns to the target Delta table on the bronze layer.

Question 9

A production team is laying out a large transactions fact table and must decide how to apply partitioning versus Z-order across a mix of low- and high-cardinality columns. Per the Databricks partitioning guidance, which TWO statements correctly describe the constraints? (Select two.)

Accepted Answer

Partitioning works well only for low or known cardinality fields such as dates, but not for high-cardinality fields like timestamps.

Accepted Answer

Z-order works for high-cardinality fields and ones that may grow infinitely, such as a customer ID in an orders table.

Answer

Z-order can be applied to the very same column the table is already partitioned on, letting both techniques cluster that field together.

Answer

Z-order clustering can freely combine files across partition boundaries to co-locate high-cardinality values table-wide.

Question 10

After a dataset goes live, a platform team is tasked with updating an existing OpenSharing share in place rather than recreating it, and they must confirm which management operations the share securable supports. Which TWO actions are valid ways to update the share? (Select two.)

Accepted Answer

Enable access to a table's history data so that recipients can perform time travel queries or streaming reads on the table.

Accepted Answer

Rename a table's alias, which is the table name that is displayed to the recipient when they browse the contents of the share.

Answer

Modify the underlying data values inside the shared source tables on the recipients' behalf as part of the same share update.

Answer

Convert the existing share into a Lakehouse Federation foreign catalog so that recipients query the shared data live over JDBC.

Question 11

A retail analytics team is migrating a batch ETL job to a Lakeflow Spark Declarative Pipeline written in Python and the developer must wrap each transformation function to register it as a pipeline table. Per the DLT documentation, which option matches?

Accepted Answer

In Python, the @dlt.table decorator defines pipeline tables with optional schema specification. The published Databricks data-engineering documentation lists this as the canonical, supported behavior.

Answer

Apply the @dlt.streamingsource decorator to each function and register the resulting object via dlt.register_source('catalog.schema.table') at module import time, before the pipeline run.

Answer

Use the @pipeline.table decorator from the pyspark.pipelines package and pass a checkpoint_location keyword argument so DLT can persist incremental state between updates.

Answer

Subclass dlt.DeclarativeTable, override the build_table() method, and call super().materialize() so DLT will treat the class as the canonical declarative pipeline node.

Question 12

An engineer is troubleshooting a Spark job on Databricks where one shuffle stage is dominated by a few huge skewed tasks and many tiny ones, and wants the engine to fix this automatically. Per the Databricks Adaptive Query Execution documentation, what does AQE do?

Accepted Answer

AQE splits skewed tasks into even ones and coalesces tiny post-shuffle partitions, all at runtime

Answer

AQE turns off the shuffle and forces a nested-loop join so the skewed keys never trigger a partition

Answer

AQE rewrites the join as a cross join and samples the keys so the optimizer can presize partitions

Answer

AQE must be turned on with a config flag per query, since it is disabled by default on Databricks

Question 13

A platform team is migrating a high-write Delta Lake table off legacy change data feed because per-write CDC computation is inflating storage costs and slowing its MERGE INTO jobs, and the lead engineer is evaluating automatic change data feed as the replacement. Per the Databricks change data feed documentation, which TWO statements correctly describe automatic change data feed?

Accepted Answer

It computes row-level changes at query time rather than at write time, using row tracking for Delta Lake and row lineage for Iceberg v3.

Accepted Answer

Because changes are not computed on every write, it improves write performance and reduces storage costs versus legacy change data feed.

Answer

It eagerly computes row-level changes on every MERGE INTO and UPDATE write, which is exactly the latency cost the team wants gone.

Answer

It still requires individual per-table configuration to be enabled on each Delta table before any row-level changes can be read.

Question 14

A production team is recovering a failed multi-task Databricks job in which 3 of 12 tasks failed, and needs to minimize re-computation when restarting the run. Which repair behavior applies?

Accepted Answer

Repairing re-executes only the unsuccessful tasks plus any tasks that depend on them.

Answer

Repairing restarts every task in the job from the start to guarantee a consistent result.

Answer

Repairing skips the failed tasks entirely and immediately marks the whole job succeeded.

Answer

Repairing re-executes only the successful tasks and leaves the failed tasks untouched.

Question 15

A platform team is implementing column masks for a healthcare PHI table on Databricks and the compliance lead must explain what a column mask actually does at query time. Per the Unity Catalog security documentation, which option matches?

Accepted Answer

Column masks transform sensitive column values at query time based on user identity, replacing actual values with masked versions. The published Databricks data-engineering documentation lists this as the canonical, supported behavior.

Answer

Column masks are evaluated by the cloud KMS at decryption time and require a customer-managed key per protected column, which is the only documented way to implement PHI redaction on a Databricks Unity Catalog managed table.

Answer

Column masks rewrite the underlying Parquet files inside the Delta transaction log to physically scrub PHI before query time, which is why they cannot be combined with row filters on the same Unity Catalog managed table.

Answer

Column masks are a one-time export-time redaction applied only by Delta Sharing recipients; they have no effect on direct queries from Databricks SQL warehouses against the underlying Unity Catalog table itself.

Databricks Certified Data Engineer Professional

Sample questions

Sources

Similar exams