Data Ingestion & Acquisition: 5 DBX-DEP Questions with Full Explanations
Practice Data Ingestion & Acquisition for the Databricks Data Engineer Professional (DBX-DEP) exam with the 5 questions below. It carries about 12% of the exam score. Every option is explained, so a wrong pick still teaches something.

Databricks Data Engineer Professional (DBX-DEP) · Data Ingestion & Acquisition · 5 questions with answers and explanations
🧠Reading Data Ingestion & Acquisition questions
Ingestion questions describe a source such as Kafka, a database or files, and a change: a new topic, missed data, a schema change. How streaming checkpoints and offsets behave usually decides. Strong options work with the checkpoint (or start a new one on purpose); weaker ones change options that a running checkpoint ignores.
Three things to look for
• Starting options such as startingOffsets apply only to a new checkpoint; replaying lost data needs a new checkpoint or a separate backfill.
• Lakeflow Connect and Lakehouse Federation solve different problems: copying data in versus querying it where it lives.
• Schema changes in the source need an explicit evolution or rescue setting, not a silent drop.
Question 1 of 5
A nightly job runs INSERT INTO pg_cat.sales.daily_totals SELECT ... FROM main.gold.totals, where pg_cat is a foreign catalog on PostgreSQL. It fails, and the team wants to publish the totals to PostgreSQL consumers. What is the correct reading?
MODIFY on the foreign catalog, and granting it enables writes that are pushed down to PostgreSQLShow answer
✅ Answer: B
Difficulty: advanced · Topic: Lakehouse Federation
Federation is read-only
Lakehouse Federation (both query federation and catalog federation) gives governed read-only access to external data through Unity Catalog foreign catalogs. You can run read queries and grant read privileges, but you cannot insert, update or delete in the external system through a foreign catalog.
Pattern for publishing back
Writing to PostgreSQL needs a different mechanism, for example a JDBC write in a job that uses credentials from a secret or a UC connection, or a reverse ETL tool. Federation remains useful in the other direction, for example as the source of a CREATE TABLE AS SELECT that lands operational data in a Delta table.
Why each option works or fails
• ✗ A. MODIFY grant: no privilege makes a foreign catalog writable, because the feature is read-only for every principal.
• ✓ B. Read-only foreign catalogs (correct): federation only supports reading external data, so the insert has no valid path, and the load into PostgreSQL has to be done another way.
• ✗ C. Primary key definition: connection definitions have no primary key setting that would enable writes, and foreign tables are always read-only.
• ✗ D. Serverless requirement: compute type does not change that foreign catalogs are read-only, and serverless is not a prerequisite for writes.
Question 2 of 5
A product team wants SQL Server changes in the Delta tables within a few minutes, at the lowest cost that still meets this goal. An engineer proposes running the ingestion pipeline in continuous mode and stopping the gateway at night. Which design follows the connector's operating model?
Show answer
✅ Answer: C
Difficulty: advanced · Topic: Lakeflow Connect, SQL Server
Two components, two modes
A gateway-based SQL Server pipeline has two parts with different operating rules:
• The ingestion gateway extracts snapshots and change logs from the source. It must run on classic compute and must stay in continuous mode so that changes are not lost to the source's log retention. Stopping it also re-provisions a VM on every start, which is slow.
• The ingestion pipeline applies the data to the destination tables on serverless compute. It runs as triggered updates on a schedule; continuous mode is not supported for it.
Latency and cost
Latency is therefore set by the pipeline's schedule. Databricks suggests leaving a few minutes between runs because serverless startup takes time. A short schedule meets "within a few minutes"; running the gateway continuously is the price of not losing changes.
Why each option works or fails
• ✗ A. Stopping the gateway: the gateway must not be stopped manually, because changes can be dropped when the source log retention passes, which forces a full refresh.
• ✗ B. Continuous pipeline: the ingestion pipeline does not support continuous mode, and a gateway that starts late risks missing changes the source has already truncated.
• ✓ C. Continuous gateway plus scheduled pipeline (correct): this matches the documented model, and a schedule of a few minutes gives the needed latency while the serverless pipeline only runs for its updates.
• ✗ D. Serverless gateway: the gateway must run on classic compute, and a daily schedule would not meet a latency goal of a few minutes.
Question 3 of 5
A retailer wants to give a logistics partner a Unity Catalog table, a folder of PDF delivery notes in a volume, and a notebook explaining the data. The partner has its own Unity Catalog-enabled Databricks workspace in another cloud. Which sharing approach covers all three?
Show answer
✅ Answer: A
Difficulty: advanced · Topic: OpenSharing, Unity Catalog
Two OpenSharing modes
OpenSharing (named Delta Sharing in older docs and in some product pages) has two protocols that use the OpenSharing server built into Databricks:
• Databricks-to-Databricks shares data and AI assets with users who have a Unity Catalog-enabled workspace. It adds notebook sharing, volume sharing, model sharing, Unity Catalog governance, auditing and usage tracking on both sides. It works across accounts, clouds and regions.
• Databricks-to-Open shares tabular data with a recipient on any computing platform (Power BI, pandas, Apache Spark, and others). Tables, not notebooks, volumes or models.
Choosing
Notebooks, volumes, models (and metric views) can only be shared with the Databricks-to-Databricks flow. When the recipient is on Databricks with Unity Catalog, that mode is both richer and token free, since the sharing identifier replaces bearer tokens.
Why each option works or fails
• ✓ A. Databricks-to-Databricks (correct): the recipient has a Unity Catalog workspace, and this is the only mode that can carry volumes and notebook files in addition to tables.
• ✗ B. Open with OIDC federation: OIDC changes how the recipient authenticates, not what the protocol can carry, so the share would still be tables only.
• ✗ C. Two open shares with exported files: this moves files out of governance and loses auditing, and the recipient has a Databricks workspace for which the integrated mode exists.
• ✗ D. Open with bearer token: the open protocol shares tabular data, so volumes and notebook files cannot be included, whatever the recipient platform is.
Question 4 of 5
A silver job reads the change data feed of bronze.customers with readChangeFeed and upserts into silver.dim_customer in foreachBatch. One micro-batch spans several commits that touch customer 42: an insert, an update (rows update_preimage and update_postimage) and later a delete. A single MERGE on customer_id against this raw batch fails or gives wrong results. Which preparation of the batch is correct?
update_postimage and insert rows and ignore deletes, since the target can drop obsolete customers during OPTIMIZE_commit_timestamp, because MERGE applies matching source rows in sorted order one after another until the last one wins_commit_version per key, because the oldest change is the one that created the key in the targetupdate_preimage rows, keep the highest _commit_version per customer_id, then merge with a delete clause for _change_type = 'delete'Show answer
✅ Answer: D
Difficulty: advanced · Topic: Delta Lake, Change data feed
What the change feed contains
Change data feed rows carry the data columns plus _change_type (insert, update_preimage, update_postimage, delete), _commit_version and _commit_timestamp. A micro-batch can contain many versions of the same key.
Applying them to a target
MERGE requires that at most one source row matches each target row; multiple matches that attempt to update the same row make the statement fail because the update would be ambiguous. So reduce the batch to the latest relevant change per key before merging:
• Remove update_preimage (old values, no new information).
• Pick the latest _commit_version per key (window with row_number()).
• WHEN MATCHED AND s._change_type = 'delete' THEN DELETE, WHEN MATCHED THEN UPDATE SET *, WHEN NOT MATCHED AND s._change_type <> 'delete' THEN INSERT *.
Why each option works or fails
• ✗ A. Ignoring deletes: deleted customers would stay in the dimension forever, and OPTIMIZE compacts files, it does not remove logically current rows.
• ✗ B. Sorted multi-row batch: MERGE does not apply source rows sequentially; multiple matching rows for one target row are an error, whatever the sort.
• ✗ C. Lowest version: that keeps the oldest state and drops the later updates and the delete, so the target ends up stale.
• ✓ D. Latest change per key without preimages (correct): one source row per key makes the merge unambiguous and the delete clause removes customers whose last change was a delete.
Question 5 of 5
Auto Loader ingests XML files that look like <orders><order id="1">...</order><order id="2">...</order></orders>, each file holding many orders. With cloudFiles.format set to xml and no further options, the read fails. Which setting makes every <order> element a row?
cloudFiles.inferColumnTypes to true, which lets schema inference find the repeating elementrowTag option to order, the repeated element to be treated as a rowrowTag to orders, so that each element below the root becomes its own rowmultiLine to true, so that records spanning several lines are split at each closing tagShow answer
✅ Answer: B
Difficulty: applied · Topic: Auto Loader
XML needs a row definition
XML has no inherent record boundary. The rowTag option, which is required, names the XML element that is read as one row. For <orders><order>...</order></orders> the correct value is the repeating child, order; the root orders would produce a single row per file with an array of orders.
Related practices
• Schema inference and schema evolution work with Auto Loader for XML; unexpected content lands in the rescued data column.
• JSON and CSV have their own parser options, such as multiLine for JSON records spanning lines.
Example: for <book><page>...</page></book> the row tag is page, as in the documentation example.
Why each option works or fails
• ✗ A. inferColumnTypes: it infers column types after the rows are defined; it cannot choose which element is a row.
• ✓ B. rowTag set to the repeating element (correct): each <order> becomes a row, which is the required row definition for XML.
• ✗ C. Root element as rowTag: the root encloses all orders, so the read yields one row per file containing the whole list, not one row per order.
• ✗ D. multiLine: that option belongs to JSON and CSV parsing; XML record boundaries are set by the row tag.
🎯 Key ideas from this set
• Lakehouse Federation is read-only; land federated data into Delta tables with CTAS and publish elsewhere through a separate write path.
• Keep the gateway running continuously on classic compute and set the latency with the schedule of the serverless ingestion pipeline.
• Notebooks, volumes and models need Databricks-to-Databricks sharing; Databricks-to-Open is for tabular data to any platform.
• Reduce a change batch to the latest change per key (without preimages) before MERGE, and map deletes to a delete clause.
• XML reads need rowTag set to the repeating element that should become a row.
📊 This set at a glance
Exam: DBX-DEP · 60 questions · 120 minutes
This set: 4 advanced, 1 applied · 2 troubleshooting, 1 comparison, 1 scenario, 1 trade-off
Topics: Lakehouse Federation, Lakeflow Connect, SQL Server, OpenSharing, Unity Catalog, Delta Lake, Change data feed, Auto Loader
🔗 Related practice sets
• Can You Answer These 5 DBX-DEP Questions on Developing Code for Data Processing using Python and SQL?
• All DBX-DEP practice sets
• Every exam on this blog
📚 Beyond this set
More DBX-DEP questions: https://hyunhp.github.io/cert-bank/#exam=DBX-DEP
Study guide: https://hyunhp.github.io/cert-bank/study.html#exam=DBX-DEP
These are original practice questions, not real exam questions, and cert-bank is not affiliated with any certification vendor.
Original practice questions written for this blog, not exam dumps.
Comments
Post a Comment