Databricks Machine Learning Associate

The exam

48 scored questions · 90 min · pass mark not published · Python · March 2025 guide.

Weights

New names

Model Registry → Models in Unity Catalog · Workspace Feature Store → Feature Store in UC (FeatureEngineeringClient) · DLT → Lakeflow pipelines · Lakehouse Monitoring → data profiling · Hyperopt (unmaintained; the guide still says fmin) → Optuna or Ray Tune recommended.

MLOps and Runtime ML

Promote code (retrain per environment) by default; promote models when training is very expensive or hard to reproduce (not viable if dev can't reach prod data). Code in Git, data and models in UC, a catalog per environment, jobs as a service principal.

Runtime ML = runtime + scikit-learn, XGBoost, PyTorch, MLflow, Feature Engineering, version-tested; GPU variants add CUDA.

AutoML

automl.classify / regress / forecast(dataset, target_col=…, timeout_minutes=…); Runtime ML up to 17.x (removed in 18.0), dedicated access mode for UC data. Gives a data exploration notebook and an editable notebook per trial, all logged to MLflow. Default split 60/20/20 random; time_col splits by time.

Feature tables (UC)

fe = FeatureEngineeringClient() → create_table(name, primary_keys, schema) · write_table(name, df, mode="merge") upserts (the only mode) · create_training_set(df, feature_lookups=[FeatureLookup(table_name, lookup_key, feature_names)], label) · log_model(…, training_set=…) stores the lookups, so score_batch(model_uri=…, df=keys_df) re-joins them — that's what avoids training/serving skew. Limits: 50 tables per model, 100 on-demand features.

Offline = Delta table, batch, training. Online = Online Feature Store (Lakebase), one row by key in ms, for endpoints. UC tables beat workspace ones: one definition, UC privileges, lineage.

MLflow tracking

Experiment → runs → params, metrics, tags, artifacts, model. Autologging is on for notebook training on Runtime ML; call mlflow.autolog() elsewhere (incl. serverless). The UI shows a run's params, metrics, artifacts, logged model and signature, source notebook and dataset.

Best run: mlflow.search_runs(order_by=["metrics.val_f1 DESC"], max_results=1) → pandas; MlflowClient().search_runs(…) → Run objects.

Models in Unity Catalog

mlflow.set_registry_uri("databricks-uc"), then register_model(f"runs:/{id}/model", "main.ml.churn") or registered_model_name=. Needs a signature. USE CATALOG + USE SCHEMA + CREATE MODEL to create a model (CREATE MODEL VERSION or ownership to add versions), EXECUTE to load or serve.

Promote by repointing an alias, not retraining: set_registered_model_alias(name, "champion", 4); also delete_registered_model_alias, set_registered_model_tag. Load models:/…@champion or /4.

Explore

describe() = count, mean, stddev, min, max · summary() adds 25/50/75% · dbutils.data.summarize(df) (Public Preview) and display()'s Data Profile tab profile every column · approxQuantile.

Two continuous → scatter, Pearson stat.corr. Two categorical → stat.crosstab, chi-square. Mixed → box plot per category.

Outliers

IQR: drop outside q1 - 1.5*iqr … q3 + 1.5*iqr; no distribution assumed, safer default. Z-score: ±3 stddev, assumes normal, distorted by the outliers themselves.

Imputation

Mean = symmetric continuous · median = skewed or outliers · mode = categorical. Imputer(strategy=…) ("mean", "median", "mode") is numeric only; strings use fillna. Fit on train only.

Encoding

StringIndexer orders by frequency (commonest = 0), handleInvalid="keep" for unseen labels. OneHotEncoder emits a sparse vector, dropLast=True. Needed by linear/logistic models and NNs; avoid for high-cardinality columns feeding trees, and for ordinals.

Transform and scale

Log (log1p if zeros) for right-skewed, non-negative features or targets; pointless for features fed to trees. StandardScaler (withMean=False by default), MinMaxScaler → [0,1], RobustScaler (median/IQR): for k-means, k-NN, SVM, regularised linear models, NNs; not trees.

Algorithms and imbalance

Interpretability → linear/logistic or a shallow tree. Tabular accuracy → gradient-boosted trees. Segments → k-means (scale first). Rare events, no labels → isolation forest.

Imbalance: class weights (class_weight="balanced", scale_pos_weight, weightCol) · oversample/SMOTE (overfit risk) · undersample (loses data) · move the threshold · score with F1, recall or PR-AUC.

Estimator vs transformer

Estimator fit() → Transformer, because it learns statistics: StringIndexer, Imputer, StandardScaler, LogisticRegression. Transformer transform(): VectorAssembler, Tokenizer, fitted models. Pipeline is an estimator, PipelineModel a transformer.

Tuning

CrossValidator(numFolds=3) averages over folds: less variance, k× cost. TrainValidationSplit(trainRatio=0.75) is one hold-out: cheaper, noisier. Both refit the winner on the whole dataset as bestModel; parallelism defaults to 1.

Grid = every combination · random = usually better per unit cost · Bayesian/TPE (Optuna, Hyperopt fmin(…, algo=tpe.suggest, max_evals=…)) = usually fewer trials. Hyperopt's objective returns a loss to minimise (return -f1).

Metrics

Accuracy = balanced, equal error cost · precision = false positives cost · recall = false negatives cost · F1 = one number under imbalance · ROC-AUC = threshold-free ranking · PR-AUC = heavy imbalance · log loss = calibrated probabilities.

RMSE (target's units, punishes big misses) · MAE (robust to outliers) · R² (variance explained, can go negative). Evaluator defaults: rmse, areaUnderROC, f1.

Both errors high = underfit: add complexity or features. Train low, validation high = overfit: regularise, simplify, more data, early stopping.

Batch and streaming

One node: mlflow.pyfunc.load_model(uri), then .predict(pdf). Distributed: predict = mlflow.pyfunc.spark_udf(spark, uri), then df.withColumn("pred", predict(*cols)) — on read or readStream. Feature lookups: fe.score_batch(model_uri=uri, df=keys_df). SQL: ai_query('endpoint', request).

A Lakeflow pipeline is the streaming answer when the question also wants managed checkpoints, retries, expectations and autoscaling.

Endpoints

Custom model: subclass mlflow.pyfunc.PythonModel (load_context, predict), log_model(python_model=…, artifacts=…, input_example=…, pip_requirements=…), register in UC, create the endpoint. A dependency missing from the logged model is a common deployment error.

served_entities = {name, entity_name, entity_version, workload_size, scale_to_zero_enabled}; POST /serving-endpoints/<name>/invocations takes dataframe_split, dataframe_records or inputs. Scale to zero: no idle cost, cold start next request.

traffic_config.routes = served_model_name + traffic_percentage (e.g. 90/10): A/B tests and canaries behind one URL, rolled back by changing the percentages.

Monitoring

Inference tables log requests and responses to Delta. A monitor with the InferenceLog profile (problem type; timestamp, prediction and model-version columns) computes profile and drift metrics plus a dashboard, window over window, optionally against a baseline table, and per slice. Drift or falling quality is the retraining signal.