Databricks Certified Data Engineer Associate Practice Exam

Correct: B. Delta Lake provides ACID transactions and time travel for safe rollbacks and audit history, and Unity Catalog governs consistent, lineage-tracked access for both AI and BI — the platform combination the requirements describe.

Why not the others: manual CSV versioning (A) is fragile and has no ACID guarantees; raw object storage with ad hoc queries (C) offers no transactional rollback or governance; in-memory DataFrames (D) are ephemeral and cannot serve as an audit trail or shared source of truth.

Correct: B — Auto Loader. Auto Loader incrementally detects and processes new files as they arrive, with built-in schema inference/evolution and checkpointing so already-ingested files are not reprocessed — the standard pattern for continuous file ingestion.

Why not the others: a manual UI upload (A) does not handle continuously arriving files; re-reading the whole directory each run (C) is inefficient and reprocesses old data; exporting to a spreadsheet (D) is not ingestion. Auto Loader is purpose-built here.

Correct: B — COPY INTO. COPY INTO incrementally and idempotently loads files from cloud storage into a Delta/Unity Catalog table, tracking which files it has already ingested so re-running does not duplicate data.

Why not the others: plain INSERT INTO (A) does not track loaded files and can duplicate; re-running CTAS (C) rebuilds the whole table rather than loading incrementally; dropping and reloading (D) is wasteful and loses history. COPY INTO is the SQL incremental-load command.

Correct: B — Silver. The Medallion pattern stages raw data in bronze, cleaned and conformed data in silver, and business-level aggregates in gold. Cleaning nulls and standardizing types on bronze produces a silver table.

Why not the others: bronze (A) is the raw landing layer before cleaning; gold (C) is the curated, aggregated layer for BI/analytics built from silver; "platinum" (D) is not a standard Medallion layer. The described step yields silver.

Correct: B — a materialized view. A materialized view stores precomputed results and refreshes them, so BI reads hit cached aggregates rather than recomputing — ideal for a gold-layer serving object.

Why not the others: a standard view (A) recomputes its query on every read; a temporary view (C) exists only for the session; a CSV extract (D) is a static snapshot that goes stale. Materialized views provide refreshed, precomputed gold data.

Correct: B. A file-arrival trigger is data-driven: the job runs when new files land, which matches "run only when new data arrives" and avoids unnecessary empty runs on a clock.

Why not the others: a cron schedule (A) runs on time regardless of whether data arrived; manual runs (C) are not automated; an always-on cluster with no trigger (D) wastes compute and does not respond to file arrival. Data-driven triggers match data availability.

Correct: B. Automation Bundles (formerly Databricks Asset Bundles) define jobs, pipelines, and config as code, version them in Git, and promote consistent deployments across environments through CI/CD — exactly the modular, repeatable mechanism described.

Why not the others: notebook revision history (A) is not real version control or packaging; model aliases (C) govern ML model versions, not ETL job deployment; manual copying (D) is error-prone and unversioned. Bundles are the supported CI/CD packaging unit.

Correct: B. One task far slower than the rest with a much larger shuffle read is the classic signature of data skew. Adaptive query execution with skew-join handling splits the oversized partition at runtime; salting the key is the manual alternative.

Why not the others: adding executors (A) does not help a single skewed task that one machine must process; collapsing to one shuffle partition (C) worsens parallelism; a network outage (D) would not produce this specific per-task skew signature. Diagnose skew from the stage metrics.

Correct: B — a managed table. Unity Catalog manages both the metadata and the data files for a managed table, so dropping it removes the underlying files. Knowing managed vs. external behavior is a core governance objective.

Why not the others: dropping an external table (A) removes only the metadata and leaves the files in their external location; a temporary view (C) has no underlying managed data files; the behavior is not the same for all types (D), which is the whole point of the distinction.

Correct: B. Granting SELECT to the group provides read access, and a column mask or column-level security on the salary column hides that field from them — the governed, least-privilege way to meet the requirement in Unity Catalog.

Why not the others: a CSV export (A) bypasses governance and creates an uncontrolled copy; relying on people not to look (C) is not a control; granting metastore admin (D) massively over-privileges the group. Scoped GRANT plus column masking is correct.