What you will be able to do
- Name the steps AutoML automates and explain how each one shortens model development
- Explain how glassbox trial notebooks let you review, reproduce and modify AutoML results, and where each notebook is stored
- Identify the controls that limit AutoML's search: excluded frameworks, timeouts, early stopping and chronological splits
- Describe how AutoML results show up in MLflow and move on to Unity Catalog registration
Key concept
Glassbox AutoML — Databricks AutoML does more than return a winning model. For each trial it generates the source code as a notebook, so the automated search ends in ordinary code you can read, rerun and edit. That is the main difference from a black-box service.
1.What AutoML takes off your plate
Building a model by hand means a long run of repetitive work. You clean the data, pick a few candidate algorithms, tune the hyperparameters of each, compare the results and write down what you tried. Databricks AutoML automates that loop. You supply a dataset and say what kind of problem it is (classification, regression or forecasting), and AutoML works through the search for you. The documentation describes the goal as simplifying the work of applying machine learning "by automatically finding the best algorithm and hyperparameter configuration for you."
The documentation lists four things AutoML does once you have given it a dataset and a problem type:
1. Cleans and prepares your data. 2. Orchestrates distributed model training and hyperparameter tuning across multiple algorithms. Trials run in parallel on the cluster instead of one after another in your notebook. 3. Finds the best model using open-source evaluation algorithms from scikit-learn, xgboost, LightGBM, Prophet and ARIMA. 4. Presents the results, including generated source code notebooks for each trial.
You can start an experiment from a low-code UI or from the Python API. Both lead to the same automated process, so analysts who prefer clicking and engineers who prefer code get the same benefit.
| Problem type | Algorithms AutoML tries |
|---|---|
| Classification | Decision trees, Random forests, Logistic regression, XGBoost, LightGBM |
| Regression | Decision trees, Random forests, Linear regression with stochastic gradient descent, XGBoost, LightGBM |
| Forecasting (classic compute) | Prophet, Auto-ARIMA |
| Forecasting (serverless) | Prophet, Auto-ARIMA, DeepAR |
AutoML trying several algorithm families is itself an advantage. You don't have to guess up front whether a tree ensemble or a linear model fits the data better, because AutoML tries them all and ranks the results. One version detail is worth knowing: in Databricks Runtime 18.0 ML and above, AutoML is not included as a built-in library. On earlier ML runtimes it was. AutoML depends on the databricks-automl-runtime package, which is available on PyPI and also helps simplify the notebooks AutoML generates.
Checkpoint 1 of 8· Check yourself
A team wants to see whether gradient-boosted trees beat a linear model on their regression problem, without hand-tuning each one. Which AutoML capability addresses this most directly?
AutoML searches across several algorithm families (for regression: decision trees, random forests, linear regression with SGD, XGBoost and LightGBM), tunes each one and finds the best model.
“Orchestrates distributed model training and hyperparameter tuning across multiple algorithms.”Source: docs.databricks.com
Checkpoint 2 of 8· Exam question
A data scientist at a retail company is building a baseline churn prediction model and wants to see the exact preprocessing and training code Databricks used, then adjust the feature engineering steps before retraining. Which capability of Databricks AutoML directly supports this?
Correct answer: A — Each AutoML trial writes an editable glassbox notebook containing the exact preprocessing and model training code, ready to review and modify.
- A. AutoML's glassbox approach generates a full source-code notebook for every trial, including the preprocessing and modeling steps, so a practitioner can open it, understand exactly what was done, and edit the feature engineering before rerunning.
- B. AutoML does not seal off training logic behind an endpoint; the whole point of the glassbox design is that the generated notebooks remain visible and reusable rather than hidden.
- C. AutoML persists far more than hyperparameter values in the Model Registry context — the trial notebooks themselves, containing the full preprocessing and training source code, are kept and importable.
- D. AutoML does not auto-deploy models or delete the notebooks that produced them; the best trial's notebook is imported into the workspace for the user to inspect and rerun.
2.Fast baselines and a first look at your data
Databricks' MLOps guidance places AutoML in the development stage, where a data scientist is still exploring the data and looking for a model worth improving. There, the guidance says, "AutoML accelerates this process by generating baseline models for a dataset." A baseline is a reference score: once you know how well an automatically tuned model does, any hand-built model has to beat it to be worth the effort. AutoML runs and records a set of trials to produce that baseline, so you get it without writing any training code.
AutoML also helps you understand the data. It calculates summary statistics on the dataset and saves them in a data exploration notebook you can review. You can open this notebook from the training page while the experiment is still running. From Databricks Runtime 10.1 ML, AutoML also shows warnings for potential dataset problems, such as unsupported column types or high-cardinality columns, on a Warnings tab. Databricks says this check is a best effort and may not catch every issue, so treat the warnings as hints rather than a full data audit.
Checkpoint 3 of 8· Check yourself
A data scientist wants to see summary statistics for their dataset before choosing a model. What does AutoML provide for this?
AutoML saves dataset summary statistics in a notebook (the data exploration notebook). The Warnings tab is only a best-effort check, not a replacement for it.
“AutoML also calculates summary statistics on your dataset and saves this information in a notebook that you can review.”Source: docs.databricks.com
Data preparation is automated too. From Databricks Runtime 10.4 LTS ML, AutoML picks an imputation method for null values based on each column's type and content, so you don't have to write imputation code for a first pass. You can override it per column with the Impute with dropdown, but as the prediction above shows, doing so switches off semantic type detection. You can also leave columns out of training (Databricks Runtime 10.3 ML and above), but you can't remove the prediction target or the time column used to split the data. If you already keep curated features in Feature Store, AutoML can use existing feature tables to augment the input dataset.
3.Glassbox notebooks: review, reproduce, modify
AutoML's biggest advantage over a black-box service is that every result comes with its code. Each call to the Python API "trains a set of models and generates a trial notebook for each model." You can open the notebook behind the winning model, read exactly how the data was preprocessed and how the model was configured, rerun it for the same result, and edit it to keep developing. The automated search hands its work over to you as ordinary, editable code.
For classification and regression, not every notebook ends up in the same place. AutoML imports some into your workspace automatically and stores the rest only as MLflow run artifacts:
| Experiment type / notebook | Where it lands |
|---|---|
| Classification and regression: data exploration notebook | Automatically imported to your workspace |
| Classification and regression: best trial notebook | Automatically imported to your workspace |
| Classification and regression: all other trial notebooks | Saved as MLflow artifacts on DBFS; notebook_path and notebook_url are not set in TrialInfo |
| Forecasting: all trial notebooks | Automatically imported to your workspace |
To work with a trial notebook that wasn't imported, open that MLflow run and find the IPython notebook in the Artifacts section of the run page. You can import it with the AutoML experiment UI or the databricks.automl.import_notebook Python API. You can also download it, if your workspace administrators allow artifact downloads.
The generated notebooks also help with explaining a model. Classification and regression notebooks include code that uses the SHAP package to calculate Shapley values, which estimate how important each feature is to the model's predictions. Because the calculation is highly memory-intensive, it is off by default. To turn it on, go to the Feature importance section of a trial notebook, set shap_enabled = True and rerun the notebook. For MLR 11.1 and below, SHAP plots are not generated if the dataset contains a datetime column.
Checkpoint 4 of 8· Match them up
Match each classification/regression notebook to where AutoML puts it
Tap a term, then the definition that fits it.
Only the data exploration notebook and the best trial notebook are imported automatically. The other trial notebooks stay as MLflow artifacts, and you import them with the UI or import_notebook.
“Generated notebooks for other experiment trials are saved as MLflow artifacts on DBFS instead of auto-imported into your workspace.”Source: docs.databricks.com
Checkpoint 5 of 8· Exam question
A team lead wants a data scientist to quickly produce a baseline classification model for a new dataset before investing time in a hand-tuned pipeline. Which advantage of Databricks AutoML best supports this goal?
Correct answer: A — AutoML's UI and Python API accept the dataset and target column, then automatically train and evaluate several candidate models without a custom pipeline.
- A. Pointing AutoML at a dataset and target column through the low-code UI or the Python API is exactly what lets a team produce a working baseline model in minutes instead of hand-writing a training pipeline first.
- B. AutoML does not require the scientist to pre-build scikit-learn Pipeline objects; it constructs and runs the preprocessing and training pipeline itself as part of each trial.
- C. AutoML does not stop at a ranked report — it actually trains, evaluates, and produces runnable model code for each trial, not just algorithm recommendations to implement later.
- D. AutoML tries multiple algorithm families in parallel trials rather than being locked to one administrator-chosen algorithm, which is what makes automatic baseline comparison possible.
4.Keeping the search under control
Automation still leaves you in charge. When you set up an experiment you choose the dataset, the problem type, the target column and the evaluation metric. The evaluation metric is the primary metric AutoML uses to score the runs. In the Python API, these are arguments to calls such as databricks.automl.classify:
Checkpoint 6 of 8· Fill the gap
Which required keyword argument tells AutoML which column to predict?
databricks.automl.classify(
dataset: Union[pyspark.sql.DataFrame, pandas.DataFrame, pyspark.pandas.DataFrame, str],
*,
? : str,
primary_metric: str = "f1",target_col is the required keyword argument that names the column AutoML predicts. time_col and split_col only control how the data is split.
Source: docs.databricks.comThe Advanced Configuration section, and the matching API parameters, let you narrow or limit the search:
| Control | Effect |
|---|---|
| Excluded frameworks (exclude_frameworks) | Leaves the selected training frameworks out of the search (Databricks Runtime 10.4 LTS ML and above) |
| timeout_minutes | Sets how long the AutoML run may take; replaces the deprecated max_trials |
| Default stopping conditions | Forecasting stops after 120 minutes; for DBR 11.0 ML and above, the number of trials is not a stopping condition |
| Early stopping | Classification and regression stop training and tuning when the validation metric stops improving |
| Time column (time_col) | Splits the data for training, validation and testing in chronological order (classification and regression) |
Early stopping means AutoML doesn't spend compute on configurations that have stopped improving the validation metric. The time-column split keeps the evaluation honest for data where order matters, because the model is never validated on rows from before the ones it trained on. Databricks also recommends leaving the Data directory field empty. AutoML then stores the dataset as an MLflow artifact, which inherits the experiment's access permissions. A DBFS path you specify yourself does not inherit them.
Replace max_trials with timeout_minutes. max_trials was deprecated in Databricks Runtime 10.4 ML and is not supported from 11.0 ML onward, so the run length is now set by time.
Sources4
5.From experiment to MLflow runs and the registry
AutoML is built on MLflow, so the results are tracked from the start. Every trial is an MLflow run. Click a run to see the trial's parameters, metrics and tags, plus the artifacts it created, including the model. The run page also includes code snippets for making predictions with that model. When the experiment completes, you can search, filter and sort the runs table, select View notebook for best model, or open the data exploration notebook. Everything is also saved to a databricks_automl folder in the home folder of the user who ran the experiment.
From there you can move straight on to deployment. When a run completes, the top row of the runs table shows the best model according to the primary metric. Select the link in the Models column and register the model to Unity Catalog or to the workspace Model Registry. Databricks recommends Unity Catalog for the latest features. After registration you can deploy the model to a model serving endpoint. The complete UI workflow, from the first click to a running experiment, is short:
Checkpoint 7 of 8· Put it in order
Put these steps for starting an AutoML classification experiment in the UI in order
- 1.In the sidebar, select Experiments
- 2.In the Compute field, select a cluster running Databricks Runtime ML
- 3.Under Dataset, select Browse and choose the table
- 4.Click Start AutoML
- 5.In the Classification card, select Start training
- 6.Select the column to predict in the Prediction target field
You open Experiments, pick the problem-type card, then fill in the Configure AutoML experiment page from top to bottom (compute, dataset, prediction target) before starting the run.
“In the Classification card, select Start training. The Configure AutoML experiment page displays.”Source: docs.databricks.com
Checkpoint 8 of 8· Exam question
A dataset for a regression problem has several columns with missing numeric values and a column of ZIP codes stored as integers that should be treated as categorical. Which Databricks AutoML capability reduces the manual data-cleaning work needed before training?
Correct answer: A — AutoML automatically imputes missing values based on column content and detects semantic types so integer columns like ZIP codes are treated as categorical.
- A. AutoML picks an imputation method based on each column's type and content, and applies semantic type detection so an integer column like ZIP codes can be handled as categorical rather than continuous, cutting the manual prep work.
- B. AutoML does not block training on missing values — automatic imputation is one of its core preprocessing steps, so nulls do not need to be manually stripped out beforehand.
- C. AutoML performs semantic type detection precisely to catch cases like a ZIP code stored as an integer, and a user can also specify custom handling, so it is not locked to treating every numeric column as continuous.
- D. AutoML does not hand off preprocessing as a blank template; it fills in the imputation and encoding logic itself in the generated trial notebooks.
Sources4
Exam traps
Each one states something that sounds right. Open it to see what is actually true.
1.Every AutoML trial notebook for classification or regression is imported into your workspace automatically.Why is that wrong?
Only the data exploration notebook and the best trial notebook are imported automatically. The other trial notebooks are MLflow artifacts on DBFS, and you import them with the UI or databricks.automl.import_notebook.
2.AutoML notebooks compute and display SHAP feature importance by default.Why is that wrong?
The notebooks contain the SHAP code, but it doesn't run by default because it is memory-intensive. You have to set shap_enabled = True and rerun the notebook.
3.You limit an AutoML run on current runtimes by setting max_trials.Why is that wrong?
max_trials is deprecated and not supported from Databricks Runtime 11.0 ML. Set the run length with timeout_minutes.
Covered in Keeping the search under control
4.AutoML ships as a built-in library on every Databricks Runtime ML version.Why is that wrong?
From Databricks Runtime 18.0 ML, AutoML is no longer included as a built-in library.
Covered in What AutoML takes off your plate
Sources
Every claim above is drawn from one of these pages, quoted as it was written on the date shown.
- 1.
“AutoML simplifies the process of applying machine learning to your datasets by automatically finding the best algorithm and hyperparameter configuration for you.”
↩︎ What AutoML takes off your plate“Finds the best model using open source evaluation algorithms from scikit-learn, xgboost, LightGBM, Prophet, and ARIMA.”
↩︎ What AutoML takes off your plate“Shapley values are based in game theory and estimate the importance of each feature to a model's predictions.”
↩︎ Glassbox notebooks: review, reproduce, modify“For forecasting experiments, AutoML-generated notebooks are automatically imported to your workspace for all trials of your experiment.”
↩︎ Glassbox notebooks: review, reproduce, modify“AutoML also generates source code notebooks for each trial, allowing you to review, reproduce, and modify the code as needed.”
↩︎ Key concept“AutoML-generated notebooks for data exploration and the best trial in your experiment are automatically imported to your workspace.”
↩︎ Exam trap 1“Because these calculations are highly memory-intensive, the calculations are not performed by default.”
↩︎ Exam trap 2“In Databricks Runtime 18.0 ML or above, AutoML is not included as a built-in library.”
↩︎ Exam trap 4“Orchestrates distributed model training and hyperparameter tuning across multiple algorithms.”
↩︎ Checkpoint“Generated notebooks for other experiment trials are saved as MLflow artifacts on DBFS instead of auto-imported into your workspace.”
↩︎ Checkpoint - 2.
“AutoML offers a no-code UI as well as a Python API.”
↩︎ What AutoML takes off your plate - 3.
“AutoML accelerates this process by generating baseline models for a dataset.”
↩︎ Fast baselines and a first look at your data“AutoML also calculates summary statistics on your dataset and saves this information in a notebook that you can review.”
↩︎ Fast baselines and a first look at your data - 4.
“AutoML displays warnings for potential issues with the dataset, such as unsupported column types or high cardinality columns.”
↩︎ Fast baselines and a first look at your data“By default, AutoML selects an imputation method based on the column type and content.”
↩︎ Fast baselines and a first look at your data“Use existing feature tables in Feature Store to augment the original input dataset.”
↩︎ Fast baselines and a first look at your data“it stops training and tuning models if the validation metric is no longer improving.”
↩︎ Keeping the search under control“For Databricks Runtime 11.0 ML and above, the number of trials is not used as a stopping condition.”
↩︎ Keeping the search under control“Not populating this field triggers the default behavior of securely storing the dataset as an MLflow artifact.”
↩︎ Keeping the search under control“showing information about the trial run (such as parameters, metrics, and tags) and artifacts created by the run, including the model.”
↩︎ From experiment to MLflow runs and the registry“When a run completes, the top row shows the best model based on the primary metric.”
↩︎ From experiment to MLflow runs and the registry“Databricks recommends you register models to Unity Catalog for the latest features.”
↩︎ From experiment to MLflow runs and the registry“The results of each AutoML experiment, including the data exploration and training notebooks, are stored in a databricks_automl folder”
↩︎ From experiment to MLflow runs and the registry“If you specify a non-default imputation method, AutoML does not perform semantic type detection.”
↩︎ Prediction“In the Classification card, select Start training. The Configure AutoML experiment page displays.”
↩︎ Checkpoint - 5.
“Each method call trains a set of models and generates a trial notebook for each model.”
↩︎ Glassbox notebooks: review, reproduce, modify“Use timeout_minutes to control the duration of an AutoML run.”
↩︎ Exam trap 3