Two domains are 53% of the Databricks Data Engineer Associate exam

Most DEA study plans treat the five exam domains as if they weigh the same. They don't. Databricks publishes the scoring weights, and two domains carry more than half the exam.

45 questions, 90 minutes, 70% to pass. That's 32 correct.

Domain	Weight	Questions
1. Databricks Lakehouse Patform	24%	~11
2. ELT with Spark SQL and lPython	29%	~13
3. Incremental Data Processing	22%	~10
4. Production Pipelines	16%	~7
5. Data Governance	9%	~4

Domains 1 and 2 are 53% of your score. Domain 5 is four questions. Spend your time accordingly.

Domain 2: ELT with Spark SQL and Python (29%)

The heaviest domain, and the one you can't fake. You read code and say what it does.

MERGE INTO is the single highest-value thing to memorize. It shows up here as UPSERT, again in Domain 3 as the CDC pattern, and again in SCD Type 1/2 scenarios. Know the branches exactly, not roughly.

MERGE INTO target t
USING source s
ON t.id = s.id
WHEN MATCHED THEN UPDATE SET *
WHEN NOT MATCHED THEN INSERT *

Also reliably tested:

• CTAS (CREATE TABLE AS SELECT) for creating Delta tables
• Window functions: ROW_NUMBER, RANK, LAG, LEAD
• Python UDFs vs built-in Spark SQL functions. Python UDFs pay serialization overhead. There's usually a question that turns on exactly this.
• Nested JSON: the : notation, explode, from_json, schema_of_json
• The same operation written both in the DataFrame API and in Spark SQL, including spark.sql()

Domain 1: Lakehouse Platform (24%)

Concepts plus platform mechanics.

All-Purpose vs Job Clusters is guaranteed. Beyond "interactive vs production", know that Job Clusters auto-terminate when the job ends and cost less.

The rest of this domain:

• Magic commands (%sql, %python, %md), dbutils.widgets, sharing variables with %run, notebook version history
• Git Folders (formerly Repos): branch switching, the PR review flow, which file types are supported
• Delta Lake basics: ACID, time travel (DESCRIBE HISTORY, RESTORE), schema evolution (mergeSchema), and the difference between OPTIMIZE and VACUUM

Domain 3: Incremental Data Processing (22%)

This one tests judgment. Which approach fits which situation.

"COPY INTO or Auto Loader?" will be on your exam. COPY INTO for a small, bounded set of files. Auto Loader for continuous ingestion at volume.

For Auto Loader, know Directory Listing vs File Notification mode, plus cloudFiles.inferColumnTypes and cloudFiles.schemaEvolutionMode.

For Structured Streaming, know the output modes (append, complete, update), the triggers (Trigger.availableNow, processingTime), and what checkpoints actually do. Trigger.availableNow vs Trigger.once is a common pairing.

Domain 4: Production Pipelines (16%)

Lakeflow Jobs and Lakeflow Spark Declarative Pipelines.

The Expectations levels come up often enough that the mapping is worth memorizing straight:

What you want	Decorator
Log the violation, keep the row	@dlt.expect
Drop the bad row	@dlt.expect_or_drop
Stop the pipeline	@dlt.expect_or_fail

Also in scope: @dlt.table and @dlt.view, Medallion layering (Bronze, Silver, Gold), task dependencies as a DAG, retry policies, and when you need to reset a streaming checkpoint.

Domain 5: Data Governance (9%)

Four questions. Know what Unity Catalog does, not how to tune it.

• The three-level namespace, catalog.schema.table
• GRANT and REVOKE syntax, plus the trap that without USAGE you can't reach nested objects
• Automatic lineage and what the lineage graph is for
• Dynamic views with CURRENT_USER() and IS_MEMBER()

Four things about the question format

Over 70% are "choose the best option". Not "which one is correct". Several options are partly correct and you pick the best one. You'll narrow to two and sit there. That's why knowing each feature's purpose and limits beats knowing it exists.

You never write code from scratch. Questions hand you SQL or PySpark and ask what it outputs, how it behaves, or why it fails. Fill-in-the-blank does show up for MERGE INTO, Auto Loader and Lakeflow syntax, so memorize those skeletons.

Delta Lake is effectively the most-tested topic. It runs through Domain 1 as concepts, Domain 2 as MERGE INTO, and Domain 3 as CDC. OPTIMIZE, VACUUM, Z-ORDER, time travel, schema evolution.

Elimination works. One or two options are usually a clearly unrelated feature, like Unity Catalog offered where the answer is MLflow. Cut to two, then think.

Try one

A pipeline ingests CSV files that keep arriving in a cloud storage landing zone into a Delta table. The file count grows daily and is now past 100,000. You want to process only new files, efficiently. Which approach fits?

A. Schedule COPY INTO, scanning all files each run to find the new ones
B. Auto Loader (cloudFiles) with Structured Streaming, tracking processed files in checkpoints
C. Batch-read the whole zone with spark.read.csv() and find the delta with LEFT ANTI JOIN
D. Reference the CSVs as an external table and filter to the latest data in a view

Answer

B. Auto Loader detects new files and remembers what it already processed, so it doesn't slow down as the count grows. COPY INTO re-scans the file listing every run, which turns into real overhead past 100,000 files. C is expensive. D throws away ACID and time travel.

Worth noticing: A and C both produce the right result. They're wrong on efficiency, not correctness. Most questions on this exam are shaped like that.

That question comes from our practice set. We maintain 6,250 questions across all seven Databricks certifications, with an explanation on every option, and you can work through them free before deciding anything: nicheelab.com/en/databricks-exam.

One scheduling detail

The retake cooldown is 14 days and the exam costs $200, so a failed attempt costs you two weeks and another $200. Worth sitting a full timed mock before you book.

The cert is valid for 2 years. If you pass, Data Engineer Professional and Machine Learning Associate both reuse your Delta Lake, Spark and Unity Catalog knowledge, so take the next one within 2 to 3 months rather than relearning it later. We cover all seven if you're going down that path.

The long version of this breakdown, including an 8-week study plan, is free to read.