Databricks Data Engineer Associate Practice Exam

• A. metastore.schema.table
• B. workspace.catalog.table
• C. catalog.schema.table Correct answer
• D. schema.catalog.table

Option C is correct because Unity Catalog uses a three-level namespace: catalog, then schema, then table, written catalog.schema.table. Option A puts the metastore where the catalog belongs, since the metastore is the top-level container and is not part of the table name. Options B and D use the wrong levels or the wrong order.

• A. In the workspace storage layer, inside the Databricks-managed root bucket
• B. In the serverless compute plane, inside the Databricks account
• C. In the compute plane, inside the customer's cloud account Correct answer
• D. In the control plane, inside the Databricks-managed cloud account

Option C is correct because classic compute runs in the compute plane, which sits inside the customer's own cloud account. The control plane in Option D holds the Databricks-managed backend services such as the web application and job scheduler, not the cluster nodes. Option B describes serverless compute, which is a different offering from classic compute. Option A confuses the root storage bucket with where Spark processes run.

• A. Parquet data files plus a transaction log in a _delta_log folder Correct answer
• B. Avro data files plus a schema registry kept in a _schema folder
• C. JSON data files plus a set of offsets kept in a _checkpoint folder
• D. ORC data files plus a manifest of file paths in a _manifest folder

Option A is correct because a Delta Lake table is a set of Parquet data files plus a transaction log stored in the _delta_log folder. The log records every commit and is what gives Delta its ACID guarantees and time travel. Options B, C and D name file formats and folders that Delta does not use for its table layout; _checkpoint, for example, is a Structured Streaming folder.

• A. It reloads all files in the folder and duplicates the earlier rows
• B. It fails until the target table is truncated before the rerun
• C. It loads the new files and needs a checkpoint location to skip the earlier ones
• D. It loads the new files and skips the files that were already loaded Correct answer

Option D is correct because COPY INTO keeps track of the files it has already loaded, so rerunning the same statement loads only the new files and skips the earlier ones, which makes it safe to retry. Option A would only happen if that tracking did not exist. Option B is wrong because no truncation is needed, and option C is wrong because COPY INTO does not use a checkpoint location (that belongs to Auto Loader and streaming).

• A. The first attempt succeeds and the new column is added to the table with no failure
• B. The first attempt fails on the new field and the retry succeeds with the column added to the table Correct answer
• C. The first attempt fails and the retry fails too until the schema hints are edited manually
• D. The first attempt succeeds and the new field's values are stored in the _rescued_data column

Option B is correct because with the default schema evolution mode the stream stops with an error when it meets a new field, after updating the stored schema at the schema location. The retry then starts with the updated schema, and mergeSchema on the Delta write lets the table gain the new column. Option A is wrong because the first attempt does fail. Option C is wrong because the schema was already updated, so no manual edit is needed. Option D describes rescue mode, which is not the default.

• A. A scheduled COPY INTO statement with the mergeSchema copy option
• B. Structured Streaming with the file source and a hand-maintained schema
• C. Auto Loader with schema evolution enabled and a schema location Correct answer
• D. A scheduled notebook that lists the folder and reads files newer than the last run

Option C is correct because Auto Loader is built for continuous incremental file ingestion at very large file counts and supports schema evolution through a schema location. Option A can handle new columns, but COPY INTO is better suited to smaller file counts. Option B requires a hand-maintained schema, so new columns would break it. Option D leaves file tracking and schema handling to custom code.

• A. Cache the transactions DataFrame before running the join
• B. Wrap the lookup DataFrame in broadcast() in the join call Correct answer
• C. Repartition the transactions DataFrame by the join key before joining
• D. Change the inner join to a cross join and filter on the key afterward

Option B is correct because broadcasting the 5 MB lookup DataFrame sends a copy to every executor, so the join happens locally and the 2 TB DataFrame is not shuffled. Option A only caches data and does not remove the shuffle. Option C itself triggers a shuffle of the large DataFrame. Option D makes the join more expensive.

• A. Use row_number over a window partitioned by event_id and ordered by updated_at descending, then keep row 1 Correct answer
• B. Call distinct() on the full DataFrame to remove the repeated event_id values
• C. Sort by updated_at descending and then call dropDuplicates on event_id
• D. Group by event_id and select first(updated_at) next to the other columns

Option A is correct because row_number over a window partitioned by event_id and ordered by updated_at descending gives a deterministic ranking, and keeping row 1 returns the latest row per event. Option B only removes rows that are identical in every column, so several rows per event_id remain. Option C is unreliable because dropDuplicates does not guarantee which row it keeps after a sort. Option D picks values per column independently, so the result does not correspond to the latest row.

• A. A Delta table rebuilt by a daily job that appends the new aggregates
• B. A standard view that aggregates the silver table at query time
• C. A streaming table that reads the silver table and is refreshed daily
• D. A materialized view defined over the silver table and refreshed daily Correct answer

Option D is correct because a materialized view stores the precomputed aggregate, so dashboard queries are fast, and each refresh recomputes the result to reflect the updates and deletes made by MERGE. Option A only appends and so cannot reflect updates or deletes. Option B recomputes at every query, which is not fast. Option C is wrong because a streaming table expects an append-only source, and MERGE updates and deletes on the silver table break that requirement.

• A. The condition task is retried two times before the branch tasks are evaluated
• B. The condition task runs with a null row count and sends the run down load_skip
• C. The condition task evaluates to false, so load_skip runs and the run succeeds
• D. The condition task and both branch tasks do not run, and the run ends as failed Correct answer

Option D is correct because the condition task uses the default run-if setting, which requires all upstream tasks to succeed. Once extract has failed on all attempts, the condition task is not run and neither are the branch tasks, and the run ends as failed. Option A is wrong because retries apply to the failing task, not to downstream tasks. Options B and C are wrong because the condition task never runs, so it produces no value and no branch is chosen.

• A. A scheduled trigger that runs every five minutes
• B. A continuous trigger that restarts the job after each run completes
• C. A table update trigger that watches the upstream table Correct answer
• D. A file arrival trigger on the upstream table's storage location

Option C is correct because a table update trigger watches a Unity Catalog table and starts the job after the table is updated, and it does not start a run when the table has not changed. Option A runs on a fixed clock whether or not the table changed. Option B reruns the job continuously regardless of the upstream table. Option D reacts to new files in a storage location, not to updates of a table.

• A. The third task is marked skipped and task four runs against the output that task two produced
• B. The job restarts at task one and discards the completed results of tasks one and two before the retry
• C. The third task starts again from its beginning while tasks one and two keep their completed results Correct answer
• D. The third task resumes from its last completed notebook cell by reading a saved execution checkpoint

Option C is correct because a task retry reruns only the failed task from its beginning, and the completed upstream tasks keep their results. The job does not restart at task one (B), the task is not marked skipped (A), and a notebook does not resume from a saved cell checkpoint (D).

• A. prod_catalog, because a target override outranks a value from the command line
• B. dev_catalog, because the default value is used when the bundle is deployed
• C. The deploy fails, because the variable is already set in the prod target
• D. qa_catalog, because the --var value outranks the target override and the default Correct answer

Option D is correct because a value passed with --var on the command line has the highest precedence, above a target override and above the default. The deploy therefore uses qa_catalog. Option A reverses the order of the first two. Option B ignores both the override and the command-line value. Option C is wrong because a target can override a variable and a command-line value can still replace it without error.

• A. databricks bundle deploy -t prod, then databricks bundle run -t prod nightly_job Correct answer
• B. databricks bundle sync -t prod, then databricks bundle run -t prod nightly_job
• C. databricks bundle init -t prod, then databricks bundle run -t prod nightly_job
• D. databricks bundle validate -t prod, then databricks bundle run -t prod nightly_job

Option A is correct because databricks bundle deploy -t prod deploys the bundle to the prod target, and databricks bundle run -t prod nightly_job then starts the deployed job. Option D validates the bundle configuration, which checks it for errors but deploys nothing, so the run command has no deployed job to start. Option C uses init, which creates a new bundle from a template. Option B uses sync, which copies local files to the workspace but does not create the job.

• A. Clone the repository into a second Git folder and commit the new transformation straight to main there
• B. Pin the production job to the latest commit of main and keep editing on main in the Git folder
• C. Edit the notebook on main in the Git folder, then commit it and hold the push until the next morning
• D. Create a feature branch from main in the Git folder, then commit and push the changes to that branch Correct answer

Option D is correct because a feature branch keeps the new transformation off main, so the nightly job that runs main is unaffected until a reviewed merge. Committing straight to main in another Git folder (A) or editing main (B, C) puts the change on the branch the job uses.

• A. Add a random salt column to the grouping key, then aggregate the partial results by customer_id Correct answer
• B. Increase spark.sql.shuffle.partitions from 200 to 2000 so the large key is split across more tasks
• C. Call repartition on customer_id before the groupBy so the rows are spread evenly across tasks
• D. Choose worker types with more memory per core so the large task stops spilling to disk

Option A is correct because all rows with the same customer_id are shuffled to one task, so one hot key creates one slow task. Salting the key first spreads that customer's rows over many tasks, and a second aggregation by customer_id combines the partial results. Adaptive query execution skew handling applies to joins, not to a groupBy aggregation, so it does not split this stage. Option B fails because a larger number of shuffle partitions still sends the one key to one partition. Option C fails for the same reason, since repartition by customer_id hashes every row of that key to the same task. Option D may reduce the spill but leaves the one slow task in place.

• A. The stage contains more tasks than the cluster has available CPU cores
• B. All executors report the same count of completed tasks for the stage
• C. The stage takes longer to complete than the stage that ran just before it
• D. The maximum task duration and input size are many times the median values Correct answer

Option D is correct because skew means a few tasks get far more data than the rest, so the slowest task and the largest input in the stage's task statistics sit many times above the median. Option A describes a stage that needs several waves of tasks, which is normal. Option B and option C do not point at uneven data: equal task counts per executor say nothing about task size, and a stage can be slower than its predecessor simply because it does more work.

• A. df.select('region', 'total') on the same DataFrame with 200 partitions
• B. df.groupBy('region').sum('total') on the same DataFrame Correct answer
• C. df.filter(col('status') == 'ok') on a DataFrame with 200 partitions
• D. df.withColumn('total', col('qty') * col('price')) on the same DataFrame

Option B is correct because an aggregation by key must bring all rows with the same key together, which requires a shuffle and ends the stage. Select, filter and withColumn are narrow transformations: each output partition depends on one input partition, so they run inside the same stage without moving data between executors.

• A. An error, since Unity Catalog blocks a new table on a path an earlier table used
• B. All the original rows, since dropping an external table leaves its files in place Correct answer
• C. No rows, since dropping the table deleted the data files at its storage location
• D. No rows, since UNDROP TABLE must first restore the dropped table within its recovery window

Option B is correct because dropping an external table removes only its metadata from Unity Catalog and leaves the data files in the storage location. A new external table created over the same location and format reads those files and returns the original rows. Option C describes a managed table drop, where Unity Catalog manages the files. Option A is wrong because the drop removes the table's registration, so the path can be used again. Option D is wrong because the files were never deleted, so no restore step is needed.

• A. Grant SELECT on the table to analysts and revoke it from all other groups
• B. Set a row filter function on the table that returns false for users outside analysts
• C. Create a dynamic view over the table that masks email and point all queries at it
• D. Attach a column mask function to email that checks is_account_group_member for analysts Correct answer

Option D is correct because a column mask function attached to the email column returns the full value when is_account_group_member shows the user is in analysts and a masked value otherwise, while users keep querying the same table name. Option A controls access to the whole table instead of masking one column. Option B filters rows, not column values. Option C forces everyone to switch to a different object name.