
Databricks Certified Machine Learning Associate checks that you can do everyday machine learning work on Databricks: explore and clean data, engineer features, train and tune models, track them with MLflow, govern them in Unity Catalog, and deploy them for batch, streaming or real-time scoring. It is a good certification for data science and ML freshers, and for data engineers who want to move into ML, because Databricks is widely used by product companies and GCCs in Hyderabad and Bengaluru.
This guide follows the official Databricks exam page and exam guide: exam facts, the four sections with weights, the code patterns you need, a 4-week plan and 15 original practice questions.
Quick facts (Databricks, November 2026)
| Item | Detail |
|---|---|
| Questions | 48 scored multiple-choice and multiple-selection questions, plus up to 10 unscored |
| Time | 90 minutes (time for unscored items is included) |
| Fee | US$200 plus applicable taxes; a retake costs the same and is allowed after 14 days |
| Delivery | Online proctored or test centre; no test aids |
| Languages | English, Japanese, Brazilian Portuguese, Korean |
| Prerequisites | None; about six months of hands-on ML on Databricks recommended |
| Validity | 2 years; recertify by passing the current exam |
Read the official exam page and download the exam guide two weeks before your exam, as Databricks advises.
Exam sections and weights
| Section | Weight | Objectives (from the exam guide) |
|---|---|---|
| 1. Databricks Machine Learning | 38% | MLOps best practices; ML runtimes; AutoML; feature tables in Unity Catalog (create, write, train and score with them; online vs offline); MLflow runs, logging and the UI; registering models in the Unity Catalog registry; promoting code vs models; model tags; champion and challenger aliases |
| 2. Data Processing | 19% | Summary statistics; removing outliers by standard deviation or IQR; visualising and comparing features; imputing with mean, median or mode; one-hot encoding and when not to use it; log transformations |
| 3. Model Development | 31% | Choosing an algorithm; handling imbalance; estimators vs transformers; training pipelines; Hyperopt fmin; grid, random and Bayesian search; parallel tuning; cross-validation; classification and regression metrics; bias-variance trade-off |
| 4. Model Deployment | 12% | Batch, real-time and streaming serving; deploying custom models to endpoints; batch inference with pandas; streaming inference with Lakeflow Declarative Pipelines (Delta Live Tables); querying endpoints; splitting traffic between models |
Code patterns to know
import mlflow
from mlflow import MlflowClient
from databricks.feature_engineering import FeatureEngineeringClient
mlflow.set_registry_uri("databricks-uc") # use the Unity Catalog registry
fe = FeatureEngineeringClient()
# Create a feature table governed by Unity Catalog
fe.create_table(name="main.ml.customer_features",
primary_keys=["customer_id"], df=features_df,
description="Customer features for churn")
# Train and log a run
with mlflow.start_run() as run:
model.fit(X_train, y_train)
mlflow.log_metric("f1", f1)
mlflow.sklearn.log_model(model, "model",
registered_model_name="main.ml.churn_model")
# Find the best run and promote a version with an alias
client = MlflowClient()
best = client.search_runs([exp_id], order_by=["metrics.f1 DESC"], max_results=1)[0]
client.set_registered_model_alias("main.ml.churn_model", "champion", 3)
from hyperopt import fmin, tpe, hp, STATUS_OK
space = {"max_depth": hp.quniform("max_depth", 2, 10, 1)}
def objective(params):
score = train_and_score(int(params["max_depth"]))
return {"loss": -score, "status": STATUS_OK}
best = fmin(fn=objective, space=space, algo=tpe.suggest, max_evals=30)
Formulas worth remembering
- Models trained in grid search with cross-validation = number of parameter combinations × number of folds (plus one final refit if the tool refits on all data).
- IQR outliers: below Q1 − 1.5 × IQR or above Q3 + 1.5 × IQR.
- Log-transformed target: exponentiate predictions before computing RMSE or MAE in the original units.
4-week study plan
| Week | Focus | Lab (Databricks Free Edition) |
|---|---|---|
| 1 | Section 2: summary statistics, outliers, imputation, encoding, log transforms | Clean a churn or housing dataset in a notebook with Spark and pandas |
| 2 | Section 3: algorithms, pipelines, cross-validation, Hyperopt, metrics | Train a pipeline, tune it with Hyperopt and compare F1 vs AUC |
| 3 | Section 1: AutoML, feature tables, MLflow tracking, Unity Catalog registry, aliases | Run AutoML; log your own model; register it and set champion and challenger aliases |
| 4 | Section 4: batch, streaming and real-time serving, traffic splitting; review | Score a table in batch; deploy a serving endpoint and query it; book the exam |
15 practice questions with answers
These are original questions written for this guide, not exam questions.
1. Why register models in Unity Catalog instead of the workspace registry?
Unity Catalog gives central governance, access control, lineage and sharing across workspaces.
2. How do you mark the model version that production should use?
Set an alias such as "champion" on that version.
3. What does AutoML give you besides a trained model?
Generated notebooks for each trial and a data exploration notebook that you can review and edit.
4. What is the difference between offline and online feature tables?
Offline tables serve training and batch scoring; online tables serve low-latency lookups for real-time inference.
5. A feature has a skewed distribution with outliers. Mean or median for imputation?
Median, because it is less affected by outliers.
6. When is one-hot encoding a poor choice?
When a feature has very many categories, because it creates a huge sparse matrix; tree models can also handle categories in other ways.
7. When does a log transformation help?
When values are positive and strongly right-skewed, such as income or prices.
8. In Spark ML, what is the difference between a transformer and an estimator?
A transformer converts a DataFrame with transform(); an estimator learns from data with fit() and returns a transformer (a model).
9. Grid search with 4 × 3 parameter values and 5-fold cross-validation trains how many models?
60 during the search (12 × 5), plus a final model if refit on all data.
10. What is one downside of cross-validation compared with a single validation split?
It costs more compute time, because the model is trained once per fold.
11. Which Hyperopt algorithm performs Bayesian-style search?
TPE (tpe.suggest).
12. A fraud model must catch as many fraud cases as possible. Which metric matters most?
Recall.
13. Which regression metric is easiest to explain in the target's own units?
MAE (RMSE is also in the same units but penalises large errors more).
14. Predictions are needed once a day for 10 million rows. Which serving approach?
Batch inference.
15. How do you test a new model on 10% of live traffic?
Serve both models on one endpoint and split traffic 90/10.
After this certification
- Move into GenAI with the Databricks Generative AI Engineer Associate, or data with the Data Engineer Associate.
- Compare cloud ML exams: AWS MLA-C02 and Google Cloud ML Engineer.
- Prepare for interviews with Databricks interview questions.
- Learn it hands-on in our AI, GenAI and agentic AI course and Multi-Cloud Data, AI and DevOps course; see live ML and data jobs.
Cloud Soft Solutions in Ameerpet, Hyderabad teaches Python, machine learning, Databricks and generative AI with real-time projects. See the APEX program and all courses.
📞 Call +91 96660 19191 · WhatsApp · Flat No. 513, 5th Floor, Nilagiri Block, Aditya Enclave, Ameerpet, Hyderabad
Frequently Asked Questions
How many questions are on the Databricks ML Associate exam?
48 scored multiple-choice and multiple-selection questions, plus up to 10 unscored items, in 90 minutes.
How much does the Databricks ML Associate exam cost?
US$200 plus applicable taxes. A retake costs the same and is allowed after 14 days.
Which section has the most weight?
Databricks Machine Learning at 38%, covering AutoML, feature tables, MLflow and the Unity Catalog model registry.
How long is the certification valid?
2 years. You recertify by passing the current exam.
Does the exam need prior certification?
No. About six months of hands-on ML on Databricks is recommended.
Can I practise for free?
Yes. Databricks Free Edition is enough for the hands-on labs in this guide.



