cert-bank

Guides

이 글을 한국어로 읽기

8 Mistakes That Repeat Across Every Databricks Exam, and How to Spot Each One in the Question

We have written 1,960 practice questions for eight Databricks exams, from the Data Engineer and Data Analyst Associates to the Machine Learning Professional, Generative AI Engineer and Context Engineer exams. Each question comes with an explanation of why every wrong option is wrong. Reading those explanations side by side, the same few mistakes keep coming back, whether the topic is Delta tables, Spark joins or RAG agents. This article names eight of them, with the clue in the question that gives each one away.

Common mistakes · Databricks certifications (DBX-DEA, DBX-DEP, DBX-DAA, DBX-MLA, DBX-MLP, DBX-GAI, DBX-CEA, DBX-ASD)

8 Mistakes That Repeat Across Every Databricks Exam, and How to Spot Each One in the Question

1. Choosing the old name, or the feature that is not finished

Databricks renames products often, and exam guides lag behind the documentation. 34 of our Databricks explanations have to point out a "formerly" name. At the same time, an option built on a Preview, Beta or Legacy feature is not the answer in this question bank, whose answer keys use generally available features only.

You may readCurrent name or replacement
Mosaic AI Vector SearchDatabricks AI Search
Lakehouse MonitoringData profiling
Databricks Asset BundlesDeclarative Automation Bundles
Genie spaces · AI GatewayGenie Agents · Unity Gateway
Model stages in the workspace registryModels in Unity Catalog with aliases such as Champion
Hyperopt · trigger(once=True)Optuna or Ray Tune · trigger(availableNow=True)

The clue: an option that works but relies on something marked Preview, Beta, Legacy, deprecated or "no longer maintained". Recognise the old names, and pick the current, generally available path.

2. Granting the table and forgetting its parents

Reading a Unity Catalog table needs three grants: USE CATALOG on the catalog, USE SCHEMA on the schema and SELECT on the table. 63 of our Databricks explanations come back to these usage privileges, in data engineering, analytics, ML and GenAI questions alike. The trap has many forms. A steward with MANAGE still cannot read: MANAGE does not include SELECT, and using it also needs the usage grants on the parents. BROWSE shows metadata but does not let anyone query. A model's creator needs USE CATALOG and USE SCHEMA as well as CREATE MODEL.

The clue: "already has SELECT (or MANAGE) but still gets a permission error". Grants inherit downward, never upward.

Beginner tip: when a permission question shows an object, write its full path as three boxes (catalog, schema, object) and put the needed privilege in each box before you look at the options. For a table that is USE CATALOG, USE SCHEMA, SELECT. For a function it is USE CATALOG, USE SCHEMA, EXECUTE. The missing box is usually the answer.

Reading a Unity Catalog object needs a grant at every level
Catalog
USE CATALOG
Schema
USE SCHEMA
Object
Table: SELECT · Function: EXECUTE · New model: CREATE MODEL
Fill one box per level before reading the options: the empty box is usually the answer, since grants inherit downward, never upward.

3. Treating a label or a sentence as a lock

Several features signal intent without enforcing anything, and the exam tests whether you know which ones:

  • The certified and deprecated values of system.certification_status are labels. A deprecated table can still be read until privileges change or the table is dropped.
  • A prompt rule such as "only use approved documents" is not access control. Our DBX-CEA explanations repeat it: constrain retrieval at the grant boundary and in the corpus, not with wording.
  • A tag on a model version records a validation result but does not stop a pipeline.
  • A budget threshold set to "send alert" keeps access open. Only "block usage" refuses requests.

The clue: the stem needs something to be prevented, and an option only describes, tags or asks.

4. Buying capacity for a design problem

"Add workers", "use a bigger warehouse", "raise the partition count", "use a larger context window": these options are tempting and usually wrong, because the bottleneck is not where the capacity goes.

  • Skew in a GROUP BY: one hot key still hashes to one partition, however many partitions or cores you add. Salting or pre-aggregation spreads it (DBX-DEP).
  • Gaps in the jobs timeline with a busy driver and idle workers: more workers cannot help, because they are waiting for the driver (DBX-DEP).
  • A serving endpoint whose model inference latency doubled while the queue is empty: the time is spent in the model or its feature lookups, so more replicas do nothing. High queue time with flat inference latency is the case that needs more capacity (DBX-MLP).
  • A larger embedding window matters only if texts overflow the smaller one (DBX-GAI). A larger context window only delays context problems (DBX-CEA).

The clue: a metric that shows one component overloaded while the others idle.

Going deeper: why AQE fixes skewed joins but not skewed aggregations

Adaptive Query Execution re-plans at shuffle boundaries using the real partition sizes. For sort-merge and shuffle hash joins, it can split an oversized partition into several tasks, because each piece of the big side can be joined separately against the matching rows of the other side. A plain aggregation has no second side to replicate. All rows for one key must end up in one place to produce one result, so splitting the partition would give partial answers. The documented remedy is to do that split yourself: aggregate by key plus a random salt first, then aggregate the small partial results by key. The hot key's work is spread over many tasks, and the second stage handles only a few rows per key.

5. Asking for "the latest" instead of passing the exact reference

When one step produces something the next step needs, the right design passes the exact reference. A training task hands the registered model URI to validation through task values. Asking the registry for the newest version breaks as soon as two runs overlap. Consumers load @Champion, and promotion moves the alias, so batch jobs and endpoints follow without code changes. In Lakeflow Jobs, a For each task passes large inputs through a lookup task and a task value reference.

The clue: words like "latest", "most recent" or "highest version" in an option about a pipeline that can run concurrently.

6. Fixing the wrong half

Many scenarios have two possible failure points, and the options fix both. Find which half failed before you choose:

  • RAG: did retrieval return the right passages? If not, no prompt change helps. If it did and the answer still invents claims, fix generation.
  • Spark out-of-memory: a collect over spark.driver.maxResultSize is a driver problem. A Java heap error in join tasks is an executor problem, which more shuffle partitions can relieve.
  • Slow SQL: compare the duration columns in system.query.history. Time spent queued at capacity calls for compute changes, not query tuning.

The clue: the stem gives evidence (a trace, a log line, a UI tile). Use it to place the fault before reading the fixes.

7. Comparing two things measured differently

A comparison proves something only if everything except the thing you changed stayed the same. Our DBX-GAI explanations insist on comparing models on your own evaluation set with the same prompt and scorers, not on numbers measured elsewhere. DBX-MLP applies the same rule to models: judge the challenger and the champion on the same fresh holdout, not against a metric stored months ago. DBX-CEA adds the spread: one lucky run is not a reliable configuration.

The clue: an option that compares a new score with an old one, a leaderboard, or a run on different data.

8. Forgetting that nothing has run yet

Spark transformations are lazy, and the DBX-ASD and data engineering exams test it from several sides. cache() stores nothing until the first action. Every action starts at least one job of its own. A temporary view is just a name for a plan, so each query against it recomputes the plan unless the DataFrame is cached. Errors that depend on the data appear only when an action runs.

The clue: code that ends without count, collect, show or a write, or an option that says something "loads immediately".

Using this list

When you miss a practice question, label it with one of these eight patterns before you read the explanation. After a few sets, one or two labels will dominate. Those are your real weak spots, and they usually cut across domains.

Practise

Practise in the question bank: every question explains each option, right and wrong. Start with the domain you just studied, then take a timed mock exam.

Counts come from the explanations in this question bank. Product names, Preview status and defaults change often; check each exam's official guide and the current Databricks documentation.