Databricks Data Engineer Professional exam: 2026 study guide

The Databricks Data Engineer Professional exam has two outlines online, and only one is current. The exam guide Databricks published on 9 October 2026 has 60 scored questions in 9 sections. The certification page that links to it still lists the earlier outline: 59 questions in 10 sections, with different weights. Study from the guide. The rest is the same on both: 120 minutes, $200 plus tax, online with a proctor or at a test centre, no aids, valid for 2 years.

This guide is for someone who already runs pipelines on Databricks and wants to pass on the first attempt. Databricks recommends a year of that work before you sit it, and the objectives show why: many begin with "choose" or "compare", between designs that would all run, for a given cost, latency or access pattern.

What moved in the October guide

Python and SQL pipeline code is still the largest part, at 23%. Below it, the questions shifted. Cost and performance rose from 13% to 15%. Ingestion grew from 7% to 12%, because sharing and federation no longer have a section of their own and their objectives now sit under ingestion. The old transformation, cleansing and quality section became Data Manipulation at 12%, which adds semi-structured data and AI functions to the joins, windows and expectations it had. Security, governance and modelling each lost a point or two.

Try itWhere the Professional exam's questions moved

Hours you have to study

60hours

Python and SQL code for data processing13.8 h23% was 22%

Cost and performance9 h15% was 13%

Data ingestion and acquisitionnow with sharing and federation7.2 h12% was 7%

Data manipulation7.2 h12% new

Monitoring and alerting6 h10% was 10%

Debugging and deploying6 h10% was 10%

Security and compliance4.8 h8% was 10%

Data governance3 h5% was 7%

Data modeling3 h5% was 6%

Transformation, cleansing and qualitygone0% was 10%

Data sharing and federationgone0% was 5%

The sum matters more than the order. Code, cost and the 2 enlarged sections add up to 62% of the exam. If your study plan came from a course written for the 10-section outline, it probably spends too long on sharing and too little on ingestion from databases and message buses.

The guide's languages line also differs. It lists English. The certification page lists English, Japanese, Brazilian Portuguese and Korean. If you plan to sit it in a language other than English, check with Databricks which outline that version follows.

The names older material gets wrong

The guide uses the product names of late 2026, and a question will not explain an old one. Map these before you read anything written before this year:

One trap sits inside this table. AUTO CDC has the same syntax as APPLY CHANGES, but it runs in Lakeflow Pipelines on Databricks, and open-source Apache Spark Declarative Pipelines does not support it. Objective 1.8 asks you to choose between Structured Streaming and Spark Declarative Pipelines, so know which features come with which. Our rename table has every change with its date.

Topics that were not on the exam a year ago

These appear by name in the October objectives and are thin or missing in older practice tests:

• CDC from databases with Lakeflow Connect, from SQL Server, MySQL and PostgreSQL, and CDC into Iceberg as well as Delta targets.
• VARIANT for semi-structured data, with parse_json and variant_get.
• ai_query inside a pipeline, to classify or enrich rows with a model.
• System tables for cost, audit and monitoring: billing, compute, access and Lakeflow.
• Alerts on business metrics, data quality, cost and query health. The guide calls them Lakehouse alerts; the documentation calls them Databricks SQL alerts, and an alert can run as a task in a job so later tasks branch on its result.
• ABAC policies with governed tags, for row filters and column masks across many tables at once.
• Erasure that satisfies GDPR, using Delta Lake and Unity Catalog features.
• Unity Catalog metric views for governed metric definitions.

The erasure objective is a good example of how the exam asks. On a table with deletion vectors, DELETE only marks rows as deleted, so the data is still in the files. REORG TABLE ... APPLY (PURGE) rewrites those files, and VACUUM removes the old ones once they pass the retention threshold, 7 days by default. An answer that stops at DELETE, or runs VACUUM straight after it, leaves the customer's rows in storage.

A study plan of 4 builds

Reading the documentation tells you what a feature is. The exam asks when to use it, which you learn by building. These 4 builds touch every section of the guide. Do them in a workspace you can use for this, such as Databricks Free Edition, a no-cost workspace for learning.

1. A CDC pipeline (sections 1, 2 and 3). Land change records as files and load them with Auto Loader into a Lakeflow Pipeline. Write the customer table with AUTO CDC twice, once as SCD type 1 and once with stored_as_scd_type set to 2, and compare the rows. Add expectations that warn, drop and fail. Put one JSON column in as VARIANT and query it. Then answer objective 1.6 for yourself: which of your tables should be streaming tables and which materialized views, and why. Our lessons on change data capture and exactly-once streaming cover the theory.

2. A job you deploy and repair (sections 1 and 8). Run the pipeline from a Lakeflow Job with an If/Else task and a For Each loop. Keep the code in a Git folder and deploy it with a Declarative Automation Bundle, with a dev and a prod target. Write 2 unit tests with assertDataFrameEqual. Then break a task with a wrong parameter, and fix the run with a repair and a parameter override instead of starting again.

3. A table you make cheaper and watch (sections 4, 5 and 9). Take a large table, convert it to liquid clustering, and compare query profiles before and after: data skipped, join strategy, shuffles. Read the Spark UI for the same job. Find what the job cost in system.billing.usage, where usage_metadata.job_id names the job on serverless and job compute but not on all-purpose compute. Set an alert on a quality check. Build one metric view over the gold table.

4. Access you can defend (sections 6 and 7). Tag the personal columns, mask them with one ABAC policy, and grant read access on a catalog to a group, with the USE CATALOG and USE SCHEMA it also needs. Then erase one customer completely, as described above. Our Unity Catalog privileges and personal data lessons go through both.

If you have no database to read from, run build 1 from files, then read how a Lakeflow Connect database connector is set up: objective 2.3 asks you to configure CDC ingestion from SQL Server, MySQL and PostgreSQL.

Practice, then the exam

Once the builds are done, test yourself against the October objectives. Our free Data Engineer Professional sample exam has 11 questions of our own across all 9 sections, with the reason behind every answer. Daily Bytes has a Data Engineer Professional prep course, and its Android app has 3 full mock exams, timed and scored by section, with Daily Bytes Plus. Use a score by section to choose what to rebuild, not to memorise answers: the real questions are different.

Databricks does not publish the pass mark, and a failed attempt can be retaken after 14 days at the full fee. 60 questions in 120 minutes is 2 minutes each. When a question gives a constraint ("lowest cost", "without rewriting data", "for tables created later"), read it before the options: it is usually what separates 2 designs that would both work.

If you have not decided between this exam and another, all 8 Databricks certifications compared shows which one to take first for the work you do.