DBX-MLA Study Guide: Two Exams in One, a Model's Life in Unity Catalog and the Arithmetic of Machine Learning
The Databricks Certified Machine Learning Associate exam mixes two kinds of question that need two different ways of studying. One kind asks where things live on the platform: which client logs a model, which name it is registered under, which alias a batch job loads, which store serves a feature. The other kind is plain machine learning, and many of those questions want a number you calculate: how many models a cross-validation trains, what the recall is, where the outlier fence sits. This guide splits the exam along that line and gives each half its own method.
Study guide · Databricks Certified Machine Learning Associate (DBX-MLA)

The exam
48 scored questions in 90 minutes, multiple choice and multiple selection, plus up to 10 unscored items. The pass mark is not published; check the official exam page. Four sections: Databricks Machine Learning 38%, Model Development 31%, Data Processing 19%, Model Deployment 12%. In practice that is about 18, 15, 9 and 6 questions.
Read the weights as two halves. Databricks Machine Learning and Model Deployment (50%) are about the platform: MLflow, Unity Catalog, Feature Store, AutoML, Model Serving. Data Processing and Model Development (50%) are mostly general ML practice, expressed with Spark ML and Python. Of the 240 DBX-MLA questions in this question bank, 162 (about two thirds) put a function, class, parameter or column name in code format, 37 are "Select TWO" multiple-selection questions, and MLflow (39 questions), Unity Catalog (38) and Feature Store (34) are the platform names that come up most often.
Half one: follow a model through the platform
Platform questions are easier when you can tell the life of one model in order. Learn this sequence until you can say it without notes.
| Step | What happens | Trap the bank keeps testing |
|---|---|---|
| Compute | Databricks Runtime ML: common ML and deep learning libraries pre-installed and tested together. | Upgrading or removing a pre-installed library breaks AutoML runs. |
| Features | A Unity Catalog Delta table with a primary key is a feature table. FeatureLookup plus create_training_set builds training data. For new projects the Feature Store docs now recommend Feature Views, which define features declaratively. | Scaling the DataFrame after load_df: that step is not repeated at inference. |
| Tracking | MLflow runs log parameters, metrics and artifacts in an experiment; search_runs finds the best run programmatically. | Parameters are inputs, metrics are results; mixing them up in the options. |
| Registry | Register in Unity Catalog under catalog.schema.model. Needs USE CATALOG, USE SCHEMA and CREATE MODEL. | Answers that move a model to "Staging" or "Production" describe the legacy Workspace Model Registry. |
| Promotion | Aliases such as Champion and Challenger point to versions; workloads load models:/prod.ml.churn@Champion. | Changing code to deploy a new version, when moving the alias is enough. |
| Serving | Batch with mlflow.pyfunc.spark_udf or score_batch; real time on a Model Serving endpoint. | An endpoint for a nightly job that nobody waits on. |
Behind the table sits one recommendation the exam assumes: deploy code, not models. Training code moves from development to staging to production, and the model is trained again in each environment. And one deployment rule: choose the mode by who waits for the answer. Stored predictions read later are batch; a caller waiting on a request is real time; events scored as they arrive are streaming.
Going deeper: why aliases replaced stages
In the legacy registry a version moved through fixed stages inside one workspace, so "production" meant a label in one place. In Unity Catalog the model is a governed object in a catalog, visible to every workspace attached to the same metastore, with privileges, lineage and auditing like a table. Environments become catalogs: with deploy code, the production job trains and registers the model in the production catalog itself; when retraining is too expensive, a version is copied there with copy_model_version. Either way, an alias then marks the version to serve. An alias is a mutable name for exactly one version at a time. A batch job that loads @Champion resolves the alias at run time, so moving the alias deploys the new version at the next run without touching the job. That separation of "which version" from "which code" is why alias answers are correct and stage answers are outdated.
Online or offline features. The offline store is the feature tables themselves, used for training and batch scoring. The Online Feature Store (the docs describe it as powered by Databricks Lakebase) serves look-ups by key with low latency for real-time models. The recurring trap goes both ways: paying for an online store when nothing needs low latency, or expecting a Delta table to answer one request in milliseconds.
Half two: do the arithmetic yourself
The ML half rewards working the numbers on paper. Each line below is a calculation the bank's questions ask for. Make sure you can reproduce it, not just recognise it.
- Counting models in a grid search. 3 values of
regParamtimes 4 values ofmaxIteris 12 parameter maps. WithCrossValidator(numFolds=5)that is 12 × 5 = 60 fits, plus one final refit on all the data: 61. Read whether the question counts the refit. - Precision and recall. 1,000 rows, 100 positives, the model flags 50 of which 40 are right. Precision 40/50 = 0.80, recall 40/100 = 0.40, accuracy 0.93. High accuracy, poor model.
- IQR fences. Q1 = 20, Q3 = 50, so IQR = 30 and the fences are −25 and 95. A value of 100 is an outlier; 95 is kept.
- MAE versus RMSE. Errors of 2, 2, 2 and 10: MAE is 4, RMSE is about 5.29. RMSE is never smaller than MAE and grows when a few errors are large.
- Log targets. A model trained on ln(revenue) predicts logarithms. Apply
exp(orexpm1afterlog1p) before reporting revenue or computing an error in currency.
Beginner tip: keep a sheet of paper next to you for every practice set in Model Development and Data Processing, and write out each confusion matrix, grid size and fence before you look at the options. The wrong options in these questions are usually the numbers you get from one common slip (forgetting the refit, swapping precision and recall, using 1.5 × Q3 instead of Q3 + 1.5 × IQR). If you compute first, those options stop looking right.
The concepts behind the numbers
Beyond calculation, the ML half tests judgement in a few recurring places:
- Estimators and transformers. A
Transformeronly hastransform()(VectorAssembler); anEstimatorhasfit()and returns a model (StringIndexer,Imputer,LogisticRegression). APipelineis an Estimator; its fitted result is aPipelineModel. - Leakage. A scaler or imputer fitted on all rows before cross-validation leaks validation data into training. Put the whole
Pipelineinside the tuner. - Imbalance. With rare positives, use precision, recall, F1 or precision-recall AUC, and consider class weights (
weightCol) or resampling. - Imputation and encoding. Median for skewed numbers, mode for categories (Spark's
Imputerhandles numeric columns only), one-hot encoding for nominal features in linear models, not for ordinal ones. - Tuning. The Databricks tuning page notes that open-source Hyperopt is no longer maintained and is not included in Databricks Runtime ML after 16.4 LTS ML; it recommends Optuna or Ray Tune. Parallel trials are meant for single-machine models such as scikit-learn; a distributed Spark ML fit already uses the cluster, and Spark's
CrossValidatorevaluates parameter maps in serial unless you setparallelism.
Going deeper: why F1 punishes the weaker of precision and recall
F1 is the harmonic mean, 2PR / (P + R), not the ordinary average. The harmonic mean is pulled towards the smaller of the two numbers. With precision 0.80 and recall 0.40, the ordinary average would be 0.60, but F1 is 0.53. With precision 0.90 and recall 0.30 the average is still 0.60, yet F1 falls to 0.45. So a model cannot hide poor recall behind high precision, or the reverse. That is the property the exam relies on when it asks which single metric to compare models with on imbalanced data where both error types matter.
A study order
- MLflow tracking and Unity Catalog registration, aliases and the deploy-code pattern (the heaviest section).
- Feature tables: creating, writing, training sets, online versus offline.
- Spark ML pipelines, validation, tuning and metrics, with paper calculations.
- Data processing: summary statistics, outliers, imputation, encoding, log transforms.
- Deployment modes, endpoints and traffic splitting, then mixed sets.
AutoML is still in the exam outline, but the AutoML docs note that it is not included as a built-in library in Databricks Runtime 18.0 ML and above; learn what it produces (tuned trial runs, the best model, and generated data exploration and trial notebooks) rather than how to run it on the newest runtime.
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.
- DBX-MLA exam page · All 300 DBX-MLA questions · Timed mock exam · 19-day study plan
- By domain: Databricks Machine Learning · Data Processing · Model Development · Model Deployment
Section weights as listed in the Databricks exam guide for the Machine Learning Associate. Runtime contents, AutoML availability and Feature Store names change; check the current exam guide and documentation.