Databricks ML Professional key terms

A logging API and UI that records parameters, metrics, artifacts, and source information for each run, letting you compare experiments and reproduce results later.

A single execution of training or evaluation code that captures its own parameters, metrics, tags, and artifacts under one run ID within an experiment.

A named container that groups related MLflow runs so results from different configurations, datasets, or team members can be compared side by side.

A feature that automatically records parameters, metrics, and models for supported libraries such as scikit-learn, XGBoost, or PyTorch with a single call, reducing manual logging code.

A stored description of a model's expected input columns and output columns, including data types, used to validate requests and enable schema enforcement at serving time.

A standardized directory format packaging model weights, environment dependencies, and metadata so the artifact can be loaded, served, or registered by multiple downstream tools.

A library-specific adapter inside an MLflow model that knows how to load and predict with the saved artifact, such as pyfunc, sklearn, or torch.

The generic MLflow model flavor that wraps any Python object exposing a predict method, allowing heterogeneous models to be served through one uniform interface.

A centralized system that computes, stores, and serves feature tables with point-in-time correctness so training and inference use consistent feature values.

A Delta table registered in the feature store with a primary key and timestamp column, enabling discovery, lineage tracking, and consistent retrieval across models.

A join that retrieves the feature values that were valid at each label's timestamp, preventing future information from leaking into training data.

A mismatch between how features are computed during model training and how they are computed at inference, often caused by duplicated or divergent transformation code.

A modeling error where information unavailable at prediction time, such as target-derived or future data, contaminates training and inflates offline evaluation scores.

A resampling technique that splits data into folds, training on some and validating on others repeatedly, to estimate generalization performance more reliably than a single split.

The search over configuration settings not learned from data, such as learning rate or tree depth, often using grid, random, or Bayesian search with parallel trials.

A Python library for sequential model-based optimization that explores hyperparameter spaces efficiently and integrates with MLflow to log each trial's results.

A condition where a model memorizes training noise and performs well on training data but poorly on unseen data, detectable through validation gaps.

A technique that penalizes model complexity, for example through L1 or L2 penalties or dropout, to reduce overfitting and improve generalization to new data.

Splitting model training across multiple workers, using data parallelism or parameter servers, so large datasets or models finish in reasonable time.

A Databricks utility that launches distributed PyTorch training across cluster nodes, handling process orchestration and communication setup for the user.

A distributed deep learning framework that uses ring-allreduce to synchronize gradients across workers, historically common for multi-node training on Spark clusters.

A centralized catalog that stores registered models, tracks their versions, assigns aliases or stages, and records descriptions and approvals for governance.

An immutable snapshot of a registered model created each time a new artifact is added, identified by an incrementing integer within that model's namespace.

A mutable named pointer, such as champion or challenger, that references a specific model version so deployment code can target a stable name instead of a number.

The recorded chain connecting a deployed model version back to its training run, parameters, dataset versions, and source code for audit and debugging.

A managed HTTPS service that loads a registered model version and returns predictions, scaling automatically and supporting traffic splits between versions.

A serving configuration where endpoint capacity drops to zero after inactivity and cold-starts on the next request, lowering cost at the price of added latency.

Pre-warmed serving capacity kept ready to handle requests immediately, avoiding cold-start delays for latency-sensitive or unpredictable workloads.

Scoring large datasets offline in scheduled or triggered jobs, writing predictions to a table rather than answering individual real-time requests.

A user-defined function that operates on pandas Series or DataFrames inside Spark, enabling vectorized model scoring across partitions for batch inference.

Continuous tracking of a deployed model's inputs, outputs, and performance metrics to detect drift, degradation, or data quality problems after release.

A change in the distribution of model input features over time compared with training data, which can degrade predictions even if the underlying relationship is unchanged.

A change in the true relationship between input features and the target, meaning the same inputs now warrant different predictions and the model must be retrained.

A Delta table that automatically captures requests and responses sent to a model serving endpoint, enabling monitoring, debugging, and later retraining.

Automated pipelines that test, build, and promote model code and artifacts through environments, often triggered by repository changes and gated by evaluation thresholds.

A trigger that fires an HTTP callback when registry events occur, such as a version transition, enabling automated testing or deployment workflows.

A governance layer providing fine-grained access control, auditing, and lineage across data and models, including registered models and serving endpoints.

An ephemeral cluster created for a single workflow run and terminated afterward, commonly used for scheduled retraining or batch scoring to control cost.

An orchestrated sequence of tasks with dependencies, schedules, and retries, used to automate data preparation, training, evaluation, and deployment steps.