Databricks Machine Learning Associate key terms

A browser-based environment that hosts notebooks, folders, clusters, jobs, and experiments for an organization's users; it is the entry point for creating and running data and machine learning work.

A long-lived compute resource created in the workspace that multiple users can attach notebooks to and share; suited to interactive development and exploration rather than automated production runs.

A short-lived cluster created automatically when a Databricks job run starts and terminated when it finishes; cheaper for scheduled production workloads because no idle time is billed.

A distributed file system layer mounted into Databricks clusters that exposes cloud object storage as file paths, letting notebooks read and write data without cloud-specific credentials or SDK code.

Databricks' centralized governance layer that defines catalogs, schemas, tables, volumes, and models with fine-grained access control and lineage across workspaces.

A cluster runtime variant that bundles popular machine learning libraries such as scikit-learn, XGBoost, and PyTorch along with GPU support and MLflow, preinstalled and version-compatible.

A parameter input, such as a text box or dropdown, added to a notebook so users can pass values into cells at runtime; commonly used to parameterize jobs.

A named container for storing credentials and tokens so notebooks and jobs can reference them without hard-coding values; secret values are redacted from output and logs.

A workspace folder linked to a remote Git repository that supports cloning, committing, and branching, enabling version control and collaboration on notebook and file-based code.

A cluster setting that adds or removes worker nodes based on workload demand within a configured minimum and maximum, balancing cost against wait time for compute resources.

A table stored in Delta Lake format whose files plus transaction log provide ACID transactions, schema enforcement, and time travel over data in cloud storage.

The MLflow component that records parameters, metrics, tags, and artifacts for each run and organizes them into experiments, providing the record of model training history.

A named grouping of MLflow runs that keeps related training attempts together and serves as the unit for comparing results and setting artifact and tag locations.

A single execution of training or evaluation code that captures its own parameters, metrics, artifacts, and start and end times under an experiment.

A centralized catalog for registering models, tracking their versions, assigning aliases, and annotating them with descriptions so teams can promote and serve them consistently.

A numbered snapshot of a registered model created each time a new model artifact is added; versions are immutable and can be given aliases or descriptions.

A mutable named pointer, such as champion or challenger, assigned to a specific registered model version so deployment code can reference a stable name instead of a version number.

A scheduled or triggered workflow that runs notebooks, scripts, or pipelines on a schedule or file-arrival trigger, with retries, notifications, and dependency graphs between tasks.

A Delta table registered in Unity Catalog that contains feature columns with a primary key and timestamp key, making the features discoverable and reusable across models.

The Databricks capability for creating, storing, and serving features consistently for training and inference, backed by Unity Catalog tables and optional online stores.

A Databricks feature that automatically explores data, generates features, trains and tunes candidate models, and produces a best model with tracked experiments and generated notebooks.

A one-line configuration that automatically logs parameters, metrics, and model artifacts from supported libraries such as scikit-learn or XGBoost during training.

A record of how data and models flow from source tables through transformations to downstream tables or endpoints, used for auditing, debugging, and impact analysis.

In Spark ML, an algorithm or object that fits training data and produces a Transformer; examples include classifiers, regressors, and pipeline stages that learn parameters.

In Spark ML, an object that applies a learned or rule-based transformation to a DataFrame, adding columns such as predictions or feature vectors.

An ordered sequence of Estimators and Transformers that is fit as one unit and can be saved and reused to apply identical preprocessing and modeling to new data.

A Spark ML Transformer that combines multiple numeric columns into a single vector column, which most Spark ML estimators require as their feature input.

A Spark ML Estimator that evaluates a pipeline over a parameter grid using k-fold splits and returns the best-fitting model plus evaluation metrics.

A Spark function that runs vectorized Python code, often pandas or Arrow-based, on groups of rows for faster and more flexible feature engineering than row-at-a-time UDFs.

A Python library integrated with MLflow on Databricks for defining search spaces and running distributed hyperparameter tuning with algorithms such as Tree of Parzen Estimators or random search.

The MLflow designation of how a logged model is saved and loaded, such as sklearn or spark, which determines the runtime dependencies needed to serve it.

A description stored with a logged model of its expected input and output schema, including column names and data types, used to validate requests at serving time.

A Databricks utility that launches distributed PyTorch training across cluster nodes by wrapping a training function and managing process orchestration.

A Databricks-hosted REST API that loads one or more registered model versions and answers prediction requests with low latency, scaling compute up or down automatically.

Scoring a large dataset offline, typically by loading a model and applying it to a Delta table with Spark or a pandas UDF, then writing predictions to a table.

A Delta table that Model Serving writes request and response payloads to, enabling monitoring of endpoint traffic, data quality, and drift.

A deployment pattern where a production model (champion) and a candidate (challenger) are served or evaluated side by side so the challenger's performance can be compared before promotion.

A Databricks feature that computes statistical profiles and drift metrics over Delta tables, including inference tables, and publishes quality dashboards.

A serving endpoint backed by an online store that returns feature values for given primary keys so models can look up features at request time.