Databricks Certified Machine Learning Associate

Learning plan

About 40 hours of focused study, split into 5 phases. Pick your weekly pace and the timeline adjusts.

Estimated finish: 4 weeks

What the exam covers

Source: official exam guide. Weights can change, so check before you start.

Before you start

• Working Python with pandas and scikit-learn, and basic Spark DataFrame operations
• Core ML concepts (train/test split, overfitting, classification vs regression metrics)
• A Databricks Free Edition workspace for the labs. Databricks recommends six months of hands-on experience

1. Weeks 1–2 · 12 h

   Databricks ML platform, MLflow and Unity Catalog

   • Explain what the Databricks ML runtime includes, and how AutoML speeds up baseline models and feature selection
   • Log params, metrics, artifacts and models manually in an MLflow run, and find the best run with the MLflow Client API
   • Register a model in Unity Catalog, set and remove tags, and promote a challenger to champion with aliases
   • Explain when to promote code rather than models, and why Unity Catalog's account-level registry beats the workspace registry
   • Create a Unity Catalog feature table, write to it, train and score with feature lookups, and tell online from offline tables

   Hands-on: In Free Edition, train two scikit-learn models on a public dataset with MLflow autologging, pick the best run with MlflowClient.search_runs, register it in Unity Catalog, and give it the @champion alias. Then register a second version and swap the alias.

   Use: Exam guide (official PDF) · MLflow on Databricks · Manage model lifecycle in Unity Catalog · Feature engineering and serving

2. Week 2 · 8 h

   Data processing and feature engineering

   • Profile a Spark DataFrame with .summary() and dbutils.data.summarize, and plot categorical vs continuous features
   • Remove outliers with standard-deviation or IQR rules, and impute missing values with the mean, median or mode
   • Decide when one-hot encoding helps or hurts (tree models, high cardinality), and when to log-transform a skewed target

   Use: Databricks AutoML docs

3. Weeks 3–4 · 12 h

   Model development and tuning

   • Choose an algorithm for a scenario and handle class imbalance (resampling, class weights)
   • Explain the difference between Spark ML estimators and transformers, and build a training Pipeline
   • Tune hyperparameters with grid, random and Bayesian search (Hyperopt fmin or Optuna), and parallelise single-node tuning
   • Count how many models a grid search with k-fold cross-validation trains, and weigh cross-validation against a single train/validation split
   • Choose F1, log loss or ROC AUC for classification and RMSE, MAE or R² for regression, and exponentiate predictions of a log-transformed target before scoring

   Hands-on: Build a Spark ML Pipeline (imputer, StringIndexer, OneHotEncoder, VectorAssembler, RandomForest) on a log-transformed price target. Tune it with a 3×3 grid and 3-fold CrossValidator, then calculate RMSE on the original scale.

   Use: Optuna on Databricks · Databricks Academy – Machine Learning with Databricks courses

4. Week 4 · 4 h

   Deployment

   • Compare batch, streaming and real-time serving, and do batch inference with pandas or a Spark UDF
   • Deploy a registered model to a Model Serving endpoint, query it over REST, and split traffic between two versions
   • Explain how streaming inference runs in Lakeflow Declarative Pipelines (formerly Delta Live Tables)

   Use: Model Serving docs

5. Week 4 · 4 h

   Practice and review

   • Answer the sample questions in the exam guide and explain why each wrong option is wrong
   • Do a timed 48-question practice set in 90 minutes and aim for 80% or higher
   • Re-drill the Databricks Machine Learning domain (38%), especially MLflow Client API calls and Unity Catalog aliases

   Use: Databricks AI Prep Guide · Udemy – Databricks ML Associate practice exams

You're ready to book when…

• You can log, register, alias and serve a model in Unity Catalog without a tutorial
• You can say how many models a given grid search with cross-validation trains, and choose the right metric for a scenario
• You score 80% or higher on a timed practice set

Exam tips

• The exam has 48 scored questions in 90 minutes and may include unscored items, so answer everything.
• The exam page now names the second domain "ML Workflows" (19%), while the March 2025 PDF guide calls it "Data Processing". Check which guide version is live two weeks before your exam.