Databricks ML Professional
Free Databricks Certified Machine Learning Professional Practice Test
12 exam-style questions with full explanations — no sign-up. Score yourself, then close your gaps with the full course.
Exam fee ~$2003 exam domainsLevel Advanced2 timed practice tests in the course
✅ Free practice — no sign-up📝 Real exam-style questions💡 Detailed explanations💸 30-day money-back via Udemy
A retail analytics team needs to train 50,000 independent forecasting models, one per store, where each store's historical sales data comfortably fits into the memory of a single worker node. The team wants to parallelize this training across the cluster efficiently rather than looping sequentially on the driver. Which approach BEST meets this requirement?
The pandas Function API lets you group a Spark DataFrame by a key (e.g., store_id) and apply an arbitrary pandas/scikit-learn function to each group in parallel across executors, which is the standard pattern for many-small-models workloads that don't need distributed algorithms.
A company registers ML models in Unity Catalog. They want the model serving endpoint to always route production traffic to the current best model version without redeploying the endpoint configuration each time a new version is promoted. Which approach BEST meets this requirement?
Unity Catalog model aliases are mutable named pointers (e.g., 'champion') that can be reassigned to any version instantly via API/UI. If the serving endpoint config references the alias, moving the alias to a new version automatically changes what the endpoint serves with no endpoint redeploy or downtime.
A company wants to deploy a new fraud-detection model version but must limit risk by exposing it to only a small percentage of live traffic first, gradually increasing exposure while monitoring error rates. Which strategy fits, and how is it implemented on Databricks Model Serving?
Canary deployment on Databricks Model Serving is implemented by setting traffic_config percentages across served entities and incrementally shifting the split toward the new version while monitoring latency/error rate, which directly matches the gradual, risk-limited exposure requirement.
A data science team has a 500GB dataset that does not fit in the memory of any single node in the cluster. They need to train ONE model on the entire dataset. Which approach BEST fits this requirement?
Spark MLlib algorithms are designed to train a single model whose computation is distributed across the cluster by partitioning the data itself, which is required when the dataset exceeds single-node memory.
A data science team monitors a production feature table for distribution drift. One feature is customer age (a continuous numeric value) and another is customer segment (categorical). Which statistical test should be used to detect drift specifically in the customer age feature?
The Kolmogorov-Smirnov (KS) test compares the cumulative distributions of a continuous numeric variable between a baseline window and a current window, making it the standard test Lakehouse Monitoring uses for numeric feature drift.
A retailer needs product recommendations returned within 200ms of a webpage load request. Which serving approach should be used?
Sub-second, per-request latency requirements are the defining use case for real-time Model Serving, which hosts the model behind a low-latency REST endpoint and returns predictions synchronously for each request.
A team is running Optuna hyperparameter optimization with multiple worker processes across a Databricks cluster, and each worker needs to read and write trial results to a shared study so trials aren't duplicated or lost. Which configuration should they use?
To run distributed Optuna trials across multiple processes or machines, the study must use a persistent shared storage backend (an RDB URL such as MySQL or PostgreSQL) so all workers can read/write the same study state and avoid duplicate or lost trials.
You are configuring Lakehouse Monitoring on an inference table that logs batch predictions daily. To compute profile and drift metrics over rolling time windows (e.g., one day at a time), which parameter must you configure in the monitor definition?
The 'granularities' parameter defines the time-bucket size (e.g., '1 day', '1 hour') used to aggregate the inference table into windows for computing profile metrics and drift statistics over time.
A data science team needs to score 500 million rows nightly using a registered model and write results to a Delta table for downstream BI dashboards. There is no low-latency requirement. Which approach is most appropriate?
Large-scale, latency-insensitive scoring is the canonical use case for batch inference: wrapping the registered model with mlflow.pyfunc.spark_udf lets Spark distribute scoring across the cluster within a scheduled job, which is efficient and cost-effective for 500M rows.
During a hyperparameter tuning sweep tracked in MLflow, what is the primary purpose of using nested runs (a parent run with multiple child runs)?
Nested runs create a parent-child hierarchy in the MLflow tracking UI so that each hyperparameter trial appears as a child run under a single parent, making it easy to visually group, sort, and compare metrics/params across the whole sweep.
A team wants a production model to automatically retrain when Lakehouse Monitoring detects significant feature drift, without any manual intervention. Which design BEST implements this requirement?
Lakehouse Monitoring writes drift metrics to metric tables that can back a SQL alert; that alert can notify or trigger automation (e.g., a webhook that kicks off a Databricks Job) that runs the retraining pipeline and registers the new model version — a fully closed-loop, automated design.
A Model Serving endpoint already has two served entities named 'model-v1' and 'model-v2'. You want to canary test model-v2 by sending it 10% of traffic while model-v1 receives the remaining 90%. Which traffic_config block correctly implements this split?
Traffic splitting on a Databricks Model Serving endpoint is configured under traffic_config.routes, listing each served_model_name with its traffic_percentage; giving model-v1 90 and model-v2 10 matches the required 90/10 canary split, and the percentages sum to 100.
Ready for the real thing?
The full course has two full-length practice tests, video lessons for every exam domain, hands-on labs and detailed answer explanations.
Start my full course on Udemy →Databricks ML Professional exam — quick answers
How much does the Databricks ML Professional exam cost?
The exam fee is approximately $200 and varies by region — confirm current pricing with the certification vendor before you book.
What topics are on the exam?
It covers 3 domains: Model Development (44%), MLOps (44%), Model Deployment (12%). The full course has a dedicated chapter, lab and practice-test coverage for each.
Is this practice test really free?
Yes — all questions on this page are free with explanations and no sign-up. The paid Udemy course adds two full-length timed exams, video lessons and hands-on labs.
Will this prepare me for the real exam?
The questions mirror the real exam's style and are mapped to the official domains. This is exam-focused preparation — combine the free test with the full course's timed simulations to gauge your readiness.