The short version
This sheet is a quick review, not a course. It lists the exam facts, then the key ideas, easily confused pairs, and decision rules for each of the four sections. It does not cover everything on the exam. The Databricks exam guide is the full list, and it is the authority if anything here differs.
Exam facts at a glance
| Item | What Databricks says (read October 22, 2026) |
|---|---|
| Scored questions | 48, multiple choice or multiple selection |
| Unscored items | Up to 10 may appear, not marked |
| Time | 90 minutes |
| Passing score | Not published |
| Aids | None |
| Languages | English, Japanese, Portuguese (Brazil), Korean |
| Validity | 2 years; recertify with the current exam |
| Code | ML code in Python; non-ML data code may be SQL |
| Retake | After 14 days |
Section weights: Databricks Machine Learning 38%, ML Workflows (called Data Processing in the exam guide) 19%, Model Development 31%, Model Deployment 12%.
Section 1: Databricks Machine Learning (38%)
Key ideas
- An ML runtime comes with common ML libraries preinstalled and tested together, so you skip setup and version clashes.
- AutoML runs a set of trial models for you and gives you the code for each trial as a notebook you can edit. It speeds up model and feature selection. It does not replace your judgment about the data.
- A feature table in Unity Catalog is a table with a primary key, created through the FeatureEngineeringClient, and governed like other Unity Catalog objects. You write data to it, then build training sets and score batches with lookups from it.
- MLflow tracks runs: parameters, metrics, artifacts, and models. Use the Client API to search runs and sort by a metric to find the best one.
- Models registered in Unity Catalog use three-level names (catalog, schema, model) and aliases. An alias such as “champion” points at one model version.
Easily confused pairs
| Term A | Term B |
|---|---|
| Offline feature table: training and batch scoring | Online feature table: low-latency lookups at serving time |
| Alias: a movable pointer to a version | Tag: descriptive metadata, changes nothing about serving |
| Unity Catalog registry: aliases, governance across workspaces | Workspace registry: older, uses stages |
| Parameter: input setting of a run | Metric: measured result of a run |
| Promote code: retrain in each environment | Promote models: move a trained model artifact |
Decision rules
- “Make this version the one production reads” means set an alias, not a tag and not a stage.
- “Find the lowest error run programmatically” means search runs with an order-by on the metric.
- Choose promoting models when retraining in production is costly or the data is not available there; choose promoting code when you can retrain on production data.
Traps: thinking a tag changes what is served; mixing up where a feature table lives (Unity Catalog versus the workspace).
Section 2: ML Workflows / Data Processing (19%)
Key ideas
- Summary statistics on a Spark DataFrame come from the summary method or the dbutils data summary.
- Outliers can be removed by standard deviation (a value far from the mean) or IQR (outside a range based on the quartiles).
- Imputation fills missing values: mean for roughly symmetric numbers, median when the data is skewed or has outliers, mode for categories.
- One-hot encoding turns a nominal category into one 0/1 column per level.
- A log transformation compresses a long right tail.
Decision rules
- Skewed values with outliers: think median, not mean.
- Categories with no order and few levels: think one-hot. Many levels: expect very wide data, so think about alternatives.
- Ordered categories: one-hot discards the order, so be careful.
- Right-skewed feature or target spanning orders of magnitude: think log scale. If the target was logged, convert predictions back (exponentiate) before computing error in the original units.
Traps: a log of zero or a negative number is undefined; no imputation strategy is automatically best, so look at the distribution first.
Section 3: Model Development (31%)
Key ideas
- Imbalanced data: cost-sensitive learning (class weights), resampling, and metrics that do not hide the minority class.
- Estimators learn from data (fit, producing a model). Transformers change data (transform). A pipeline chains both.
- Tuning: Hyperopt (fmin) searches a space; random, grid, and Bayesian search differ in how they pick the next trial. Grid search tries every combination.
- Cross-validation (k folds) gives a steadier estimate than a single train-validation split, but costs k times the training.
Metrics
| Task | Common metrics | Note |
|---|---|---|
| Classification | F1, log loss, ROC AUC | Accuracy misleads on imbalanced data |
| Regression | RMSE, MAE, R-squared | RMSE punishes large errors more; MAE is easier to read |
Decision rules
- Models trained in a grid search with cross-validation: number of combinations times number of folds. Count combinations by multiplying the number of values for each hyperparameter.
- Low training error with high validation error: high variance (overfitting). Reduce complexity or add regularization.
- High error on both: high bias (underfitting). Add capacity or better features.
- Parallelizing a single-node model for tuning: think distributed trials rather than distributed training.
Traps: forgetting the fold multiplier; picking accuracy for a rare-event problem.
Section 4: Model Deployment (12%)
| Approach | Use when | Watch for |
|---|---|---|
| Batch | Scoring on a schedule, no per-request latency need | Cheapest and simplest |
| Real-time | Per-request answers from an endpoint | Always-on cost |
| Streaming | Continuously arriving data | Pipeline design, for example with Delta Live Tables |
- Batch inference with pandas means scoring a DataFrame with the loaded model.
- A custom model is deployed to a serving endpoint, then queried for real-time inference.
- Splitting traffic between served models in an endpoint uses percentages that must add up to 100.
Last-day checklist
- Read the exam guide’s outline once, and tick each bullet you can explain aloud.
- Redo the questions you missed in practice, not the ones you enjoy.
- Rehearse the counts: combinations times folds.
- Check your exam booking, ID, and system check. Our exam day guide has a calm routine.
- Stop studying early. Sleep beats a final cram.
More guides for DATABRICKS-ML-ASSOCIATE
- Databricks ML Associate Exam Format and Retakes
- Databricks ML Associate Study Plan
- Is the Databricks ML Associate Exam Hard?
- Databricks ML Associate Prerequisites
- Databricks ML Associate Renewal and Retakes
- Is Databricks ML Associate Worth It?
- Databricks ML Associate vs Professional
- GenAI vs ML Associate: Which Databricks Exam?
Next step
Go through the domains explained page for the reasoning behind each section, or test yourself with the sample questions. When you want full-length practice, try the Databricks ML Associate practice tests. Back to the exam overview.