AZ MLO key terms

A top-level Azure Machine Learning resource that groups related experiments, compute, datasets, models, and endpoints so teams can share and govern assets within one boundary.

A designated compute resource such as a compute instance, compute cluster, or attached Kubernetes cluster where training jobs or scoring services execute.

A fully managed, single-node cloud workstation used for authoring notebooks, running interactive experiments, and testing code before scaling out.

A managed pool of scalable multi-node machines that automatically provisions and tears down nodes to run training or batch scoring jobs.

A saved connection to a storage service such as blob storage or a data lake, holding credentials so jobs and users can reference data by name.

A versioned pointer to data in a datastore, recording its location, type, and metadata so experiments can reference a stable, reproducible data source.

A reusable specification of the software stack for a job or deployment, including base image, Python packages, and environment variables, enabling consistent reproducibility.

A single unit of execution submitted to Azure Machine Learning, such as a command script or pipeline run, with tracked status, logs, and outputs.

A multi-step workflow that chains jobs or components together with defined inputs and outputs, supporting reuse and scheduled retraining.

A logical grouping of runs used to compare results and track metrics across many attempts at the same modeling problem.

One execution of a job or script that records metrics, parameters, artifacts, and logs for later review and comparison.

A named, versioned entry in the model registry that points to stored model files and carries metadata such as tags and descriptions.

A central catalog that stores, versions, and organizes trained models so they can be discovered, audited, and deployed across environments.

An immutable snapshot of a registered model under a given name, incremented on each registration so deployments can pin an exact artifact.

A managed HTTPS endpoint that hosts one or more model deployments for real-time scoring, handling scaling, authentication, and traffic routing.

An instance of a model running on an endpoint with its own environment, scoring script, and resource allocation, which can receive a share of traffic.

A configuration that routes specified percentages of endpoint requests to different deployments, enabling safe canary rollouts and A/B comparison.

An endpoint that runs scoring jobs over large datasets on a schedule or on demand, writing predictions to a configured output location.

A file defining init and run functions that load a model and produce predictions from request or batch input at serving time.

A declared description of a model's expected input and output schema, used to validate requests and document the serving contract.

An open-source API and UI for logging parameters, metrics, and artifacts during runs, integrated with Azure Machine Learning for experiment comparison.

A standard directory format packaging model files, dependencies, and metadata so the model can be loaded and served by MLflow-compatible tools.

The practice of assigning immutable version identifiers to datasets so training runs can be reproduced against the exact data used.

The recorded relationship chain linking data assets, jobs, environments, and models, supporting reproducibility, debugging, and audit investigations.

A change in the distribution of input features over time relative to the training data, which can degrade model accuracy after deployment.

A configured Azure Machine Learning feature that compares recent production input data against a baseline dataset and signals when distributions diverge beyond thresholds.

A scheduled signal that evaluates a deployed model's inputs, outputs, or performance metrics and raises alerts when configured thresholds are breached.

An observability service that collects request logs, latency, error rates, and custom telemetry from online endpoints for operational troubleshooting.

A toolset that visualizes model error analysis, fairness, interpretability, and counterfactual outcomes to support responsible model assessment.

An evaluation comparing model performance across sensitive groups defined by attributes such as age or gender to detect disparate error rates.

A data field identifying a protected or demographic group, used in fairness analysis and sometimes excluded from training features.

A measure of how much each input feature contributes to a model's predictions, used to explain behavior globally or for individual cases.

The degree to which a person can understand why a model produced a given prediction, supported by explanation techniques such as SHAP or surrogate models.

The process of segmenting model errors by feature cohorts to find where performance is weakest and guide targeted improvement.

A YAML document listing channels and package dependencies used to build a reproducible Python environment for training or inference.

A capability that automatically tries combinations of algorithms, preprocessing, and hyperparameters to find strong models for a labeled dataset.

A systematic search over model configuration values using sampling methods and an early-termination policy to find better-performing settings.

The process of creating, transforming, or selecting input variables so a model can capture relevant patterns more effectively.