Free study guide · Databricks
How to pass the Databricks Certified Machine Learning Associate
What the exam covers, how to plan your study, and how to read its questions. From the CertLumen course for this exam, kept up to date with the official exam guide.
- Questions
- 48
- Time
- 90 minutes
- Passing score
- Not published by Databricks; this course targets ≥ 80% on unseen mocks
- Exam fee
- USD 200 (plus applicable taxes)
- Valid for
- 2 years (recertify by taking the current live exam)
- Delivery
- Online proctored or test center
What the exam covers
A study plan that works
The default path for data scientists and ML engineers who know Python and basic machine learning but haven't used every Databricks ML feature the exam covers. It covers all 17 lessons and all six labs in section order, with a section drill at the end of each section and short daily flashcard reviews. Sections 1 (38%) and 3 (31%) get four days each, because together they carry about 33 of the 48 scored questions. Day 12 closes the content with the deployment lab, Mock Exam 1 is on day 13 and Mock Exam 2 on day 14, and you sit the exam on day 15. Run every drill with Keep the mock-exam questions unseen ticked, so both mocks are new when you take them. Plan on about 1.5–2 hours a day.
| Day | Focus | Time |
|---|---|---|
| 1 | Kick-off: the exam, ML runtimes, AutoML and MLflow tracking | 105 min |
| 2 | Models in Unity Catalog, then the MLflow lab | 100 min |
| 3 | Feature tables: create, write, train and score | 95 min |
| 4 | MLOps strategy, the feature table lab and the m1 drill | 105 min |
| 5 | Profiling, outliers, charts and log transforms | 95 min |
| 6 | Imputation and one-hot encoding, the m2 lab and drill | 110 min |
| 7 | Algorithms, class imbalance, estimators and pipelines | 95 min |
| 8 | Hyperparameter tuning and cross-validation | 105 min |
| 9 | Evaluation metrics and the Spark ML lab | 105 min |
| 10 | The tuning lab, the m3 drill and batch inference | 115 min |
| 11 | Streaming inference and Model Serving | 95 min |
| 12 | The deployment lab, the m4 drill and a full review | 110 min |
| 13 | Mock Exam 1 | 120 min |
| 14 | Mock Exam 2 and final review | 120 min |
Exam strategy
How the exam works
The Databricks Certified Machine Learning Associate exam checks whether you can do basic machine learning on Databricks: use AutoML, Unity Catalog and MLflow, explore data and engineer features, train, tune and evaluate models, and deploy them. Most items are short scenarios that ask for the best option under a stated constraint, and many show a Python snippet. Some are multiple-selection items that tell you how many options to pick.
| Fact | Detail (from the exam page, the exam guide and the certification FAQ) |
|---|---|
| Scored questions | 48 multiple-choice and multiple-selection questions |
| Unscored items | Up to 10 may be added. They aren't marked, don't affect your score, and the extra time for them is already in the time limit |
| Time limit | 90 minutes for everything you see (up to 58 items), about 1.5 minutes per item |
| Fee | USD 200 plus applicable taxes |
| Delivery | Online proctored or at a test center. You register on Webassessor, and Kryterion delivers the exam |
| Test aids | None allowed |
| Languages | English, Japanese, Portuguese (BR) and Korean |
| Code | Machine learning code is in Python; non-ML code may be in SQL (exam page) |
| Prerequisites | None. Related training and six months of hands-on Databricks ML work are highly recommended |
| Passing score | Not published |
| Validity | 2 years. You recertify by passing the current live exam |
The exam page links the August 2026 PDF of the exam guide. Its Updates table lists March 1, 2025 for both the exam and the outline, so the objectives in this course are those of the live exam. The guide asks you to check it again two weeks before your exam date.
Section weights: where the points are
Two sections carry 69% of the score: Databricks Machine Learning (38%) and Model Development (31%). The question counts below are estimates from the weights: Databricks publishes only percentages, and exam forms vary. The exam page labels Section 2 *ML Workflows*, while the guide's outline calls it *Data Processing*. It's the same 19%.
| Section | Weight | ≈ Questions | Course section | Highest-yield topics |
|---|---|---|---|---|
| 1. Databricks Machine Learning | 38% | 18 | m1 | AutoML advantages, MLflow logging and search_runs, Unity Catalog registration, tags and aliases, feature tables (create, write, train, score), online vs offline, deploy code vs deploy models |
| 2. Data Processing | 19% | 9 | m2 | summary() vs dbutils data summaries, standard deviation vs IQR outliers, charts and comparison methods, mean/median/mode imputation, one-hot encoding, log transforms |
| 3. Model Development | 31% | 15 | m3 | Algorithm choice, class imbalance, estimators vs transformers, Hyperopt fmin, grid/random/Bayesian search, cross-validation and model counts, metrics, exponentiating log targets, bias-variance |
| 4. Model Deployment | 12% | 6 | m4 | Batch vs streaming vs real-time, pandas and spark_udf scoring, streaming inference in Lakeflow Spark Declarative Pipelines (Delta Live Tables), serving endpoints, traffic splitting |
| Total | 100% | 48 |
The guide's ten sample questions are a good preview: a feature table created with FeatureEngineeringClient.create_table, imputation chosen after examining the distribution, cost-sensitive learning for imbalance, 3 x 2 x 3 combinations x 5 folds = 90 models, a Delta Live Tables pipeline that applies the model as a Spark UDF, promotion with set_registered_model_alias, search_runs ordered ascending on RMSE, a log transform for right-skewed values, high variance for a large train-validation gap, and batch inference for nightly scoring. The course teaches each one with the same reasoning.
How to use this course
- Learn. Read the lesson, answer its self-checks and note any key point you would have missed.
- Lab. Run the lab on Databricks Free Edition. A few steps need a paid workspace (an online feature store, classic AutoML on 17.3 LTS ML) and say so; read those steps instead.
- Drill. Take the section drill untimed after each section, with Keep the mock-exam questions unseen ticked. The questions in the four mocks then stay out of the drills, so the mocks stay a true test; the drills use all the others. Read every explanation, including the notes on the options you rejected.
- Flashcards. Review in short daily sessions, and repeat missed cards until you get them right two days in a row.
- Mocks. Take Mock Exam 1 and Mock Exam 2 (48 questions, 90 minutes each) under exam conditions, never open-book. Mock Exams 3 and 4 are spares: use them for a fresh first attempt after you have worked through your misses, or before a retake.
- Review loop. Keep an error log: topic, why you missed it (knowledge gap, misread constraint or trap), the correct rule, and where it is in the course. Once you have taken every mock you plan to use, untick the box and drill every question of your weakest sections.
A repeated mock measures memory, not readiness. Only first attempts count toward the readiness criteria; use repeats to review explanations.
Readiness criteria
- ≥ 80% (39 of 48) on two mocks, each taken timed, in one sitting, on the first attempt.
- ≥ 70% in every section on those mocks: 13 of 18 for m1, 7 of 9 for m2, 11 of 15 for m3 and at least 4 of the 6 m4 questions.
- You finish a mock in ≤ 75 minutes, with time left for the flagged items.
- Without notes, you can write
search_runsfor the best run,set_registered_model_alias,create_table,write_tableandscore_batch,summary()with percentiles, the IQR filter,Imputer,StringIndexerplusOneHotEncoder,fmin,spark_udfand an endpointtraffic_config. - You can count the models in any grid search with cross-validation in under 30 seconds.
Databricks doesn't publish the passing score. Aiming for 80% on unseen mocks leaves a margin for a harder exam form, unfamiliar wording and exam-day nerves.
Time management
You have 90 minutes for 48 scored items plus up to 10 unscored ones, so plan on about 1.5 minutes per item. Recall items (a default, a method name) take 30 to 45 seconds, which leaves time for code and multiple-select items. Use three passes.
- First pass (≈ 60 min). Answer everything you can within about 90 seconds. If you're torn between two options, pick one, flag the item and move on.
- Second pass (≈ 15–20 min). Go back to the flagged items and re-read the stem's constraint words before the options.
- Final check (≈ 5 min). Make sure nothing is blank and every multiple-select item has the number of answers it asks for. Change an answer only for a concrete reason.
| Checkpoint | 48 items on screen | 58 items on screen |
|---|---|---|
| One third of the items | ≤ 28 min (item 16) | ≤ 30 min (item 20) |
| Two thirds | ≤ 56 min (item 32) | ≤ 60 min (item 39) |
| Last item | ≤ 75 min, then flagged items | ≤ 80 min, then flagged items |
An unanswered item earns nothing. On a multiple-select item, pick exactly as many options as the stem asks for. Databricks doesn't say that partial credit is given, so treat these items as all or nothing.
Multiple-selection items
- The stem tells you how many to choose, for example *(Choose two.)*, usually from five options.
- Judge each option on its own as true or false for the scenario, then pick the required number. Don't look for options that seem to belong together.
- Typical true pairs: AutoML advantages (a quick baseline, every trial tracked in MLflow), Unity Catalog benefits (access from every workspace with UC grants, lineage plus audit), offline vs online feature tables, and one-hot encoding fits (few nominal levels for linear or distance-based models).
- The distractors are usually real features in the wrong place: stages and webhooks for Unity Catalog, SHAP on by default,
SparkTrialson serverless, fractional traffic percentages.
Reading scenario questions
Most items give a few sentences of context, some of it irrelevant, and ask for the best solution. One or two constraint words usually decide the answer. Read in this order:
- Read the last sentence (the actual ask) first.
- Find the constraint words: least effort, correct results, governed, every workspace, programmatically, milliseconds, fluctuating volume, most cost-effective, no code change.
- Predict the answer before you read the options.
- Eliminate every option that breaks a constraint, then choose from what's left.
Words that point to the answer
Phrases in a question that usually point to one answer. The course has the full list.
| When the question says | Think of |
|---|---|
| pre-installed, tested ML libraries, MLflow ready, no init scripts | Databricks Runtime for Machine Learning: tick Machine learning in the create-compute form |
| quick baseline, little code, many algorithms tried, editable code for every trial | AutoML |
from databricks import automl fails on Databricks Runtime 18 LTS ML | Run AutoML on a 17.3 LTS ML cluster in Dedicated access mode |
| notebook moved to serverless, training works but no MLflow run appears | Call mlflow.autolog() before training |
| chart validation loss across epochs in the MLflow UI | mlflow.log_metric('val_loss', v, step=epoch) |
programmatically retrieve the run with the lowest validation_rmse | MlflowClient().search_runs([exp_id], order_by=['metrics.validation_rmse ASC'], max_results=1)[0] |
| one governed copy of the model for every workspace, with lineage and audit | Register in Unity Catalog with a three-level name catalog.schema.model |
| registration fails: the model has no signature metadata | Log the model again with input_example= or signature=, then register it |
| mark the new version as the one production reads, recommended approach | client.set_registered_model_alias(name, 'champion', new_version) |
record validation_status = approved on version 3 only | client.set_model_version_tag(name, '3', 'validation_status', 'approved') |
Revise from the cheat sheet
The course's cheat sheet has a page for each of these:
- Databricks ML: runtimes, AutoML, MLflow, Unity Catalog models and feature tables
- Data processing: profiling, outliers, charts, imputation, encoding and log transforms
- Model development: algorithms, imbalance, pipelines, tuning, cross-validation and metrics
- Model deployment: batch, streaming and real-time inference, Model Serving and traffic splits
Practise with the real exam's mix
270 exam-style questions, 4 timed practice exams, flashcards and an AI tutor.