DDEA
Free DDEA practice questions
10 questions from our DDEA practice tests, each with the correct answer and a written explanation. Read them with the answers shown, or switch to quiz mode and answer first.
Question 1 of 10, Databricks Intelligence Platform
A data engineering team runs a nightly ETL notebook as a scheduled job. Today it runs on an all-purpose cluster that developers also use during the day. The team wants to lower the cost of the nightly run without changing the notebook code. What should they do?
- AAttach the job to a cluster pool so idle instances left over from the daytime cluster are reused at night
- BKeep the all-purpose cluster but enable autoscaling with a lower maximum worker count overnight
- CMove the notebook to a SQL warehouse so the run is billed at the warehouse DBU rate instead
- DRun the notebook as a job task on job compute or serverless jobs compute instead of the all-purpose cluster Correct answer
Answer and explanation: Job compute (classic job clusters or serverless jobs compute) is billed at a lower DBU rate than all-purpose compute and exists only for the run, so moving the scheduled task off the interactive cluster cuts cost without code changes. Autoscaling limits still bill all-purpose rates. A SQL warehouse cannot run an arbitrary Python notebook ETL, and pools reduce startup time rather than the DBU rate charged for all-purpose usage.
Question 2 of 10, Data Ingestion and Loading
A pipeline must ingest an operational SQL Server database incrementally into Unity Catalog using a Lakeflow Connect managed connector. Which component does this connector type use to extract snapshots and change data from the source database before the ingestion pipeline applies them?
- AA JDBC connection opened from a notebook task that polls the source tables for updated timestamps
- BA partner connector from Partner Connect that replicates the database into an external location
- CA Delta Sharing recipient configured on the SQL Server side to push every table change into Databricks tables
- DAn ingestion gateway that runs on classic compute and stages extracted data in a Unity Catalog volume Correct answer
Answer and explanation: Lakeflow Connect database connectors such as SQL Server use an ingestion gateway that connects to the source, captures snapshots and change data, and stages it in Unity Catalog storage; a separate ingestion pipeline then applies the changes to target tables. Partner Connect tools are third-party alternatives, not this connector's components. A hand-built JDBC polling notebook is a custom approach, and Delta Sharing shares Delta data rather than extracting from SQL Server.
Question 3 of 10, Data Ingestion and Loading
Auditors ask the data team to record, for every bronze row, which source file it came from. The files are ingested with Auto Loader. What is the simplest way to capture this?
- ACall input_file_name() inside a Python UDF applied to each row after the write
- BQuery the Delta transaction log afterwards to match each row to its add action
- CEnable change data feed on the bronze table to record the source of every insert
- DSelect the _metadata.file_path field and store it in a column during ingestion Correct answer
Answer and explanation: File-based sources expose a hidden _metadata column whose file_path (and file name, size and modification time) can be selected during ingestion. input_file_name() is legacy and is not supported in Unity Catalog standard access mode, and using it after the write would no longer know the source. The transaction log records data files of the table, not source files, and change data feed tracks row changes, not source file paths.
Question 4 of 10, Data Transformation and Modeling
A gold aggregate of daily revenue by region must be refreshed incrementally when possible and served to a BI dashboard with precomputed results. The source is a silver table updated by batch jobs. Which Unity Catalog object fits BEST?
- AA materialized view, which stores precomputed results and can refresh incrementally on update Correct answer
- BA standard view, which stores no data and recomputes the aggregation on every dashboard query
- CA temporary view, which is precomputed per session and shared across all dashboard viewers
- DA streaming table, which is designed for append-only sources and recomputes aggregates in full
Answer and explanation: A materialized view stores the query results, and refreshes can be incremental when the engine can compute only the changes, which suits precomputed BI aggregates. A standard view computes on every query. Temporary views are session-scoped and not precomputed, and streaming tables are aimed at append-only incremental ingestion rather than maintaining aggregates over changing sources.
Question 5 of 10, Data Transformation and Modeling
A gold table is read by a dashboard that filters by country. Analysts must only see aggregated revenue, not order-level rows, and dashboards must load quickly from precomputed data. Which design meets these requirements?
- ACreate a materialized view with the aggregation and grant analysts SELECT on it but not on the silver orders table Correct answer
- BGrant analysts SELECT on the silver orders table and let the dashboard aggregate the rows on every query
- CCreate a standard view with the aggregation and grant analysts SELECT on both the view and the silver orders table directly
- DCreate a streaming table over the silver orders table with an append-only aggregation for the dashboard
Answer and explanation: A materialized view precomputes the aggregate for fast dashboards, and granting SELECT only on it keeps order-level rows hidden. Granting SELECT on the silver table exposes row-level data, and a standard view recomputes on every query. Streaming tables are meant for append-only incremental processing and do not maintain aggregates over changing data like a materialized view.
Question 6 of 10, Working with Lakeflow Jobs
A transform task in a Lakeflow Job sometimes fails because a source API returns transient errors. The team wants the task to try again up to three times, waiting a few minutes between attempts, before the run fails. Where should this be configured?
- AIn the job's maximum concurrent runs, set to 3 so three attempts run in parallel at once
- BIn the task's Run if condition, set to All done so the task reruns automatically whenever the upstream fails
- CIn the task's retry policy, setting the number of retries and the minimum interval between them Correct answer
- DIn the job's schedule, by adding three extra trigger times a few minutes apart each night
Answer and explanation: Each task has a retry policy where the number of retries and the minimum interval between attempts are set, which handles transient failures automatically. Extra schedule times launch whole new runs even after success. Concurrent runs would start parallel duplicates rather than sequential retries, and Run if conditions control whether a task starts based on its dependencies, not retries of the task itself.
Question 7 of 10, Working with Lakeflow Jobs
A job runs every 15 minutes but occasionally takes 25 minutes. The team does not want two runs to overlap and wants a triggered run to wait rather than be skipped when one is already active. Which settings achieve this?
- ASwitch to a continuous trigger so the next run always starts immediately when the current one finishes
- BSet maximum concurrent runs to 1 and enable queueing so that new runs wait for the active one Correct answer
- CSet the Run if condition of the first task to All done so it waits for the previous run
- DSet maximum concurrent runs to 2 and add a timeout of 15 minutes on every task in the job
Answer and explanation: With maximum concurrent runs at 1, a new run cannot start while one is active, and enabling queueing places the triggered run in a queue instead of skipping it. Allowing two concurrent runs causes overlap, and a 15-minute timeout would kill legitimate long runs. Run if conditions apply within a run, not across runs, and continuous mode changes the scheduling model rather than queueing scheduled runs.
Question 8 of 10, Implementing CI/CD
A team uses Declarative Automation Bundles (formerly Databricks Asset Bundles) to deploy a job. In dev, the job should write to the dev_catalog catalog; in prod, to prod_catalog, while the job definition stays the same. How should they configure this?
- ACreate two separate databricks.yml files, one per environment, each containing its own copy of the job definition
- BUse a Spark configuration in the notebook that reads the workspace URL to pick the right catalog
- CDefine a catalog variable in databricks.yml, reference it as ${var.catalog}, and override it in each target Correct answer
- DHard-code prod_catalog in the job and let developers change it locally before each dev deployment
Answer and explanation: Bundle variables are declared in the variables mapping, referenced with ${var.<name>}, and given different values per target, so the same resource definition deploys to dev and prod with environment-specific settings. Duplicate bundle files drift apart, hard-coding requires manual edits, and inferring the environment from a workspace URL is brittle and hides configuration outside the bundle.
Question 9 of 10, Troubleshooting, Monitoring, and Optimization
In the Spark UI for a slow aggregation stage, the task metrics show large values for Spill (Memory) and Spill (Disk) on most tasks, while task durations are fairly uniform. What does this indicate, and what is a reasonable first fix?
- ABroadcast joins are too large; lower spark.sql.autoBroadcastJoinThreshold so the planner forces shuffle joins
- BSkew on a single key; salt the grouping key and aggregate the partial results twice
- CThe driver is collecting results; replace collect() with a write to a Delta table
- DPartitions are too large for executor memory; raise shuffle partitions or use larger-memory workers Correct answer
Answer and explanation: Spill on most tasks with uniform durations means partitions do not fit in execution memory, so increasing the number of shuffle partitions (making each smaller) or using workers with more memory reduces spill. Skew would show one or a few outlier tasks rather than uniform spill. collect() problems appear on the driver, and the stage is an aggregation, so broadcast thresholds are not the cause.
Question 10 of 10, Governance and Security
A data engineer must create an external table over abfss://landing@corpdata.dfs.core.windows.net/orders. The CREATE TABLE statement fails with a permission error on the path. The engineer has CREATE TABLE on the target schema. Which additional object and privilege are required?
- ACREATE EXTERNAL TABLE on an external location that covers the path, which itself uses a storage credential Correct answer
- BMODIFY on the target schema, which lets any table created in it reference arbitrary cloud storage paths directly
- CREAD FILES on the storage credential directly, with no external location defined for that container
- DCREATE MANAGED STORAGE on the catalog, so the path becomes the catalog's default managed location
Answer and explanation: To create an external table, the path must fall under an external location (which pairs the path with a storage credential), and the user needs CREATE EXTERNAL TABLE on that external location plus the usual schema privileges. Granting privileges directly on storage credentials for file access is not the recommended pattern and does not grant table creation. MODIFY on a schema governs data changes, not storage paths, and CREATE MANAGED STORAGE is an external-location privilege for using a path as managed storage, not for registering an external table.