Databricks Certified Machine Learning Associate Practice Exam

What the Databricks Certified Machine Learning Associate exam covers

• Databricks Machine Learning281 questions
• ML Workflows141 questions
• Model Development229 questions
• Model Deployment89 questions

Free Databricks Certified Machine Learning Associate sample questions

A sample of 10 questions with answers and explanations. Sign up free to practice all 740.

1. Question 1Databricks Machine Learning

   A data scientist writes: with mlflow.start_run(): model.fit(X_train, y_train) mlflow.log_metric('rmse', rmse) What happens to the run if an unhandled exception is raised inside the with block?

   • AMLflow automatically marks the run FAILED, or FINISHED on success, when the block exitsCorrect
   • BIt automatically disables autologging for the remainder of the notebook session
   • CIt forces every subsequent metric call to include an explicit step value
   • DIt moves the active experiment into a new workspace folder automatically
   ✓ Correct answer: A

   mlflow.start_run() used as a context manager automatically calls mlflow.end_run() when the with block exits. If the block exits normally the run status is FINISHED, but if an exception propagates out, MLflow sets the run status to FAILED before re-raising the exception.

   Why the other options are wrong

   • BAutologging state is independent of a single run's exception handling.
   • Cstart_run does not impose any requirement on the step argument of log_metric.
   • DExceptions inside a run do not relocate or recreate the experiment.
2. Question 2Databricks Machine Learning

   What is an input_example used for when logging a model with `mlflow.sklearn.log_model()`?

   • AA backup copy of the entire training dataset
   • BA small sample of real input data saved for reference and testingCorrect
   • CA template config file for deploying to Kubernetes
   • DA synthetic random tensor used only to benchmark inference speed
   ✓ Correct answer: B

   Passing input_example logs a small, representative sample of the data the model expects, which downstream users can reference to understand the expected format, and which MLflow can also use to infer a signature automatically.

   Why the other options are wrong

   • Ainput_example is a small sample, not a full backup of the training data.
   • Cinput_example is about the model's data format, it has nothing to do with Kubernetes deployment configuration.
   • Dinput_example is meant to be a realistic sample of actual data, not a synthetic benchmarking input.
3. Question 3Databricks Machine Learning

   A team's Spark ML CrossValidator job needs to fit many candidate models across folds and hyperparameters in parallel over a large distributed dataset. Which cluster configuration best supports this?

   • AA single-node cluster with no workers and only local driver compute available
   • BA multi-node cluster with several worker nodes so Spark can parallelize fold-fitting and data processingCorrect
   • CA cluster with autotermination set to one minute to save on idle compute cost
   • DA cluster limited to the driver's local file system only, with no Delta or Spark access
   ✓ Correct answer: B

   Spark ML's CrossValidator can distribute the fitting of different folds and hyperparameter combinations across executors, so a cluster with multiple worker nodes speeds up this parallel search over a large dataset.

   Why the other options are wrong

   • AA single-node cluster with no workers has nothing extra to parallelize the fold-fitting across.
   • CA one-minute autotermination setting is about idle shutdown timing, not compute distribution for training.
   • DRestricting to the driver's local file system does not describe a cluster's compute capability for distributed training.
4. Question 4Databricks Machine Learning

   What is a key practical difference between the legacy workspace Feature Store and Feature Engineering in Unity Catalog?

   • AUnity Catalog feature tables are governed centrally and referenced consistently across workspaces, unlike workspace-local onesCorrect
   • BThe legacy workspace Feature Store supports point-in-time lookups, while Unity Catalog feature tables do not
   • CFeature Engineering in Unity Catalog cannot be used with FeatureLookup or create_training_set
   • DOnly the legacy workspace Feature Store can be written to using write_table
   ✓ Correct answer: A

   Because Unity Catalog feature tables live under a shared metastore rather than a single workspace's local storage, the same governance and naming apply consistently no matter which workspace a user connects from.

   Why the other options are wrong

   • BPoint-in-time lookups are a Feature Store / Feature Engineering capability available with both the legacy and Unity Catalog-based APIs, not exclusive to the legacy one.
   • CFeatureLookup and create_training_set are exactly the APIs used with Feature Engineering in Unity Catalog.
   • Dwrite_table is used to update feature tables in both the legacy Feature Store and Feature Engineering in Unity Catalog.
5. Question 5ML Workflows

   A data scientist inspects a Spark DataFrame loaded from a Delta table and wants to confirm which columns are nullable and what data type each column has, without triggering a job. Which call is appropriate?

   • Aspark_df.printSchema()Correct
   • Bspark_df.summary().show()
   • Cspark_df.count()
   • Dspark_df.corr('a', 'b')
   ✓ Correct answer: A

   printSchema() reads the DataFrame's schema metadata and prints a tree of column names, types, and whether each is nullable, which is a metadata-only operation and does not require scanning the underlying data.

   Why the other options are wrong

   • Bsummary() computes actual statistics over the data and triggers a Spark job, which is more than needed here.
   • Ccount() triggers a full job to tally rows and reveals nothing about column types or nullability.
   • Dcorr() is not a DataFrame method signature in this form and measures correlation, not schema.
6. Question 6ML Workflows

   What does Spark ML's Bucketizer transformer do to a continuous numeric column?

   • AIt splits the column into discrete buckets based on user-supplied split boundariesCorrect
   • BIt learns quantile-based split points automatically from the data distribution
   • CIt rescales the column into the 0 to 1 range
   • DIt removes rows that fall outside three standard deviations of the mean
   ✓ Correct answer: A

   Bucketizer takes a splits array, for example [-inf, 0, 18, 65, inf], and assigns each row to the bucket index whose boundaries contain its value. Unlike QuantileDiscretizer, Bucketizer does not learn the boundaries from the data; you must supply them directly based on domain knowledge.

   Why the other options are wrong

   • BLearning quantile-based split points automatically describes QuantileDiscretizer, not Bucketizer.
   • CRescaling into 0 to 1 describes MinMaxScaler, not Bucketizer.
   • DBucketizer assigns every value to a bucket based on the given splits; it does not remove rows as outliers.
7. Question 7Model Development

   Given `model = indexer.fit(df)` where indexer is a StringIndexer, how do you apply the learned index mapping to `df`?

   • ACall `model.fit(df)` a second time.
   • BCall `model.transform(df)`.Correct
   • CCall `model.predict(df)`.
   • DCall `df.transform(model)`.
   ✓ Correct answer: B

   Once fit, `model` is a Transformer (StringIndexerModel). Calling `model.transform(df)` applies the learned string-to-index mapping and returns a new DataFrame with the indexed column added.

   Why the other options are wrong

   • Amodel is already fit; calling `.fit()` again would attempt to re-learn the mapping rather than apply it.
   • CSpark ML Transformers use `.transform()`, not `.predict()`, to apply their logic to a DataFrame.
   • DThe transform call is made on the model object with the DataFrame as its argument, not the other way around.
8. Question 8Model Development

   The F1 score is best described as which of the following?

   • AThe arithmetic mean of precision and recall
   • BThe harmonic mean of precision and recallCorrect
   • CThe product of precision and recall
   • DThe proportion of variance explained by the classifier
   ✓ Correct answer: B

   F1 = 2 * (precision * recall) / (precision + recall). Using the harmonic mean means F1 stays low if either precision or recall is low, unlike a simple arithmetic average.

   Why the other options are wrong

   • AAn arithmetic mean would be (precision + recall) / 2, which does not penalize a low value in either metric as strongly as F1 does.
   • CF1 is not simply the raw product of precision and recall; that quantity is not scaled the same way.
   • DVariance explained describes R2, a regression concept unrelated to F1.
9. Question 9Model DevelopmentSelect all that apply

   Which TWO hp functions sample on a logarithmic scale, appropriate for parameters spanning several orders of magnitude such as learning rate? (Choose TWO)

   • Ahp.loguniformCorrect
   • Bhp.uniform
   • Chp.qloguniformCorrect
   • Dhp.choice
   • Ehp.quniform
   ✓ Correct answer: A, C

   Both give small and large values comparable sampling density on a multiplicative scale, unlike linear distributions.

   Why the other options are wrong

   • Bhp.uniform samples linearly, under-weighting small values relative to large ones for multiplicative parameters.
   • Dhp.choice samples from a discrete list of values, unrelated to log-scaled numeric sampling.
   • Ehp.quniform quantizes a linear (non-log) uniform distribution, not a logarithmic one.
10. Question 10Model Deployment

    Which scenario is the poorest fit for streaming inference?

    • AScoring sensor readings as they continuously arrive in a Kafka topic
    • BScoring transactions as they land in a Delta table in near real time
    • CGenerating a one-time report from a static, unchanging CSV exportCorrect
    • DFlagging anomalies in a continuously updating event log
    ✓ Correct answer: C

    Streaming inference is valuable when data keeps arriving over time and needs to be scored continuously. A one-time report from a fixed, unchanging file is better served by a single batch job.

    Why the other options are wrong

    • AContinuously arriving Kafka data is a classic streaming inference scenario.
    • BNear-real-time scoring of newly landed Delta rows is exactly what streaming inference targets.
    • DContinuously updating logs are naturally suited to a streaming pipeline.

Who this Databricks Certified Machine Learning Associate practice exam is for

This practice set is for anyone preparing for the Databricks Certified Machine Learning Associate exam at the intermediate level - from first-time candidates building a foundation to experienced Data practitioners doing a final review before test day. If you learn best by working through realistic questions and reading why each answer is right or wrong, it is built for you.

How to use this Databricks Certified Machine Learning Associate practice exam

1. Start with the free sample questions above to gauge your current baseline.
2. Read the full explanation on every question, including why each wrong option is wrong.
3. Track your weak domains and focus your study where you are losing the most marks.
4. Once you are scoring consistently well, take a timed, full-length mock exam.
5. Use your readiness score to decide when you are ready to book the real Databricks Certified Machine Learning Associate exam.

Related Data resources

• Databricks Certified Machine Learning Associate study guideKey concepts
• Data practice examsAll Data
• Certification pathWhere this fits
• Certification exam guides & tipsBlog
• Plans & pricingFree & paid
• Databricks Certified Data Engineer Associate practice examRelated
• SnowPro Advanced practice examRelated
• SnowPro Core (COF-C02) practice examRelated

Databricks Certified Machine Learning Associate practice exam FAQ