The short version
This sheet is a review aid, not a substitute for study. It lists the exam facts, then the ideas that scenario questions lean on in each of the three domains: Model Development (44%), MLOps (44%), and Model Deployment (12%). It does not cover everything on the exam. Use it to spot gaps, then go back to the Databricks exam guide and the platform documentation for depth.
Exam facts at a glance
| Item | What Databricks publishes |
|---|---|
| Scored questions | 59, multiple choice |
| Unscored items | Up to 10 may appear and are not marked |
| Time | 120 minutes |
| Language | English |
| Delivery | Online proctored or test center |
| Test aids | None |
| Passing score | Not published |
| Retake wait | 14 days |
| Validity | 2 years, renewal by the full current exam |
| Guide date | September 30, 2025 |
Details are on the exam format and renewal and retakes pages.
Domain 1: Model Development (44%)
Spark ML
- A pipeline chains stages. An estimator learns from data and produces a model. A transformer applies a fixed change.
- Typical order for mixed columns: index categories, encode them, assemble a feature vector, then fit the model.
- Evaluate with an evaluator that matches the task: regression metrics for a number, classification metrics for a label.
- Tune with a parameter grid plus cross-validation; every extra grid point multiplies the fits.
- Decision rule: if the data is large and training must scale, think SparkML. If it fits on one machine and you need a library SparkML lacks, think single node.
Scaling and tuning
- Data parallelism splits the data across workers with the same model. Model parallelism splits the model itself.
- Vertical scaling is a bigger machine. Horizontal scaling is more machines.
- Optuna tuning can be spread across the cluster and logged to MLflow. Ray is a separate distributed framework you can also run on Spark.
- Pandas Function APIs train or score one model per group. Each group must fit in one worker’s memory.
- Decision rule: “one model per customer or store” means a grouped pandas Function API. “One big model on huge data” means distributed training.
Advanced MLflow
- Nested runs: a parent run for the whole search and child runs for each configuration; the final model links back to the parent.
- Log metrics with a step when a value changes over time, and log files as artifacts.
- A custom model class (PyFunc) can hold preprocessing logic plus artifacts it loads at prediction time.
- Trap: an artifact used in prediction that was never logged with the model.
Feature Store concepts
- Point-in-time correctness means each training row sees feature values as they were at that row’s event time, not later values.
- Offline feature tables feed training and batch scoring. Online tables serve low-latency lookups. On-demand features are computed at request time from the request’s inputs.
- Streaming ingestion keeps features fresh without batch recomputation.
- Decision rule: if the stem mentions leakage or “future data,” think timestamp lookups.
Domain 2: MLOps (44%)
Lifecycle and environments
- The deploy-code idea: promote the code through dev, test, and prod, and train the model in each environment, instead of copying model artifacts across.
- In Unity Catalog, a model version is pointed to by an alias. Aliases replace the older stage names.
- Databricks Asset Bundles describe jobs, experiments, registered models, and serving endpoints as code. The usual order is validate, deploy, run.
- Decision rule: “version-controlled, reproducible, multiple environments” points to a bundle.
Testing
- Unit tests check one function with known inputs and outputs. Integration tests check that stages work together: features, training, evaluation, deployment, inference.
- Change something early in the pipeline and more downstream tests must be re-run. A change only to a training setting leaves feature engineering untouched.
- Trap: running slow end-to-end tests on every small edit.
Automated retraining
- Triggers: a drift alert or a performance drop.
- Choose the selection metric from the business cost. If decisions depend on probability thresholds, a metric that scores probability quality is a better fit than one that only checks hard labels.
Lakehouse Monitoring
| Term | Meaning |
|---|---|
| Snapshot profile | Whole table at one point in time |
| Time series profile | Data with a timestamp, compared across windows |
| Inference profile | Model inputs, predictions, and labels over time |
| Profile metrics table | Statistics per column and slice |
| Drift metrics table | Change versus a baseline or the previous window |
- A baseline is a fixed reference. Consecutive windows compare each period with the one before it.
- Custom metrics, slicing by a column, and alerts on thresholds extend the basics.
- Endpoint health (latency, request rate, error rate, CPU, memory) is separate from data drift.
- Trap: reading metric tables before the monitor has refreshed.
Domain 3: Model Deployment (12%)
- Blue-green: two full environments, traffic switches at once. Canary: a small share of traffic goes to the new model first and grows while you watch.
- For a high-traffic, business-critical path, think gradual rollout, not an instant 100% switch.
- Model Serving can host several served entities behind one endpoint with a traffic split.
- A custom PyFunc model is registered in Unity Catalog with its artifacts, then queried with REST or the MLflow Deployments SDK.
- Trap: hand-building HTTP calls when an SDK call handles authentication and payload shape.
Easily confused pairs
- Estimator is for learning. Transformer is for applying.
- Data parallelism is for big data. Model parallelism is for big models.
- Baseline comparison is against a fixed reference. Consecutive-window comparison is against the prior period.
- Alias is a movable pointer. A version number is a fixed identity.
- Offline table is for training. Online table is for low-latency serving.
- Blue-green is a switch. Canary is a gradual shift.
- Unit test is one function. Integration test is a chain of stages.
Last-day checklist
- Re-read the topic list in the current exam guide and check the date.
- Say the confusable pairs above out loud, without looking.
- Review the questions you got wrong in practice, not the ones you got right.
- Confirm your booking, ID, and system check if you test online.
- Stop new material the evening before and sleep.
More guides for DATABRICKS-MLP
- Databricks ML Professional Study Plan
- Databricks ML Professional Domains Explained
- Is the Databricks ML Professional Exam Hard?
- Databricks ML Professional Prerequisites
- Is Databricks ML Professional Worth It?
- Databricks ML Associate vs Professional
Next step
Test what you remember with the sample questions, or work through full sets in the Databricks ML Professional practice tests. The full topic list is in the Databricks exam page.
Frequently asked questions
Is this everything I need?
No. It is a review aid. The exam guide and the platform documentation are the authority on what can be asked.
Does it contain real exam questions?
No. The decision rules are our summaries of concepts, not recalled questions.