Azure DP-100 flashcards

32 cards. Read the term and the definition here. The deck at the top of this page is optional.

Azure ML workspace

Top-level Azure resource that groups compute, datastores, data assets, environments, models, endpoints, and experiments, and holds shared configuration including linked Key Vault and container registry.

Zero minimum cluster nodes

Setting min_nodes to 0 lets a compute cluster scale down to zero nodes when idle, so no virtual machine cost accrues while no jobs are running.

Serverless compute

Fully managed compute that Azure ML allocates per job or notebook session, removing the need to create and maintain a dedicated compute cluster or instance.

Credential-less datastore access

A datastore configuration that authenticates using the workspace's managed identity instead of storing account keys or SAS tokens inside the workspace.

MLTable data asset

A data asset type that points to tabular files and carries a schema definition, so training scripts can read delimited or parquet data without custom parsing code.

Azure ML environment

A reusable, versioned specification bundling a base Docker image, Conda or pip dependencies, and environment variables so training and scoring run reproducibly across computes.

Workspace managed identity

A system- or user-assigned Entra ID identity that the workspace uses to authenticate to attached storage accounts, Key Vault, and container registries without stored secrets.

Private endpoint workspace

A workspace configured with private endpoints and public network access disabled, so clients and dependent services communicate only over an Azure virtual network.

MLflow autolog

A single call that automatically records parameters, metrics, and model artifacts for supported frameworks such as scikit-learn, PyTorch, and XGBoost into the tracking server.

Sweep sampling algorithms

Hyperparameter tuning supports random, grid, and Bayesian sampling; random suits large search spaces, grid exhaustively tests discrete values, and Bayesian uses prior trial results.

Bayesian sampling concurrency

Bayesian sampling evaluates trials sequentially because each trial is informed by earlier results, so it does not support parallel trials or most early-termination policies.

Median stopping policy

An early-termination policy that cancels a trial when its primary metric falls below the median of running averages from all prior trials at the same interval.

AutoML primary metric

The metric AutoML optimizes, such as accuracy, AUC_weighted, or normalized_root_mean_squared_error, paired with a goal of maximize or minimize.

Population stability index

A data drift metric comparing the distribution of a feature or prediction between a baseline dataset and a recent dataset; larger values indicate greater shift.

SHAP feature importance

Shapley-value explanations that quantify each feature's contribution to individual predictions and aggregate into global importance rankings inside the responsible AI dashboard.

Feature set materialization

The process that computes feature set data and writes it to an offline or online store so training jobs and real-time inference read consistent features.

Scoring script init and run

An inference script defines init() to load the model once at startup and run(raw_data) to transform request data into a JSON-serializable prediction payload.

Traffic split across deployments

An online endpoint routes a percentage of requests to each deployment; shifting traffic gradually enables blue-green and A/B testing behind one stable scoring URL.

Online endpoint autoscaling

With scale type default, a deployment adjusts instance count to keep average CPU or request utilization near target_utilization_percentage; manual mode fixes a constant instance_count.

Batch endpoint scoring function

Batch deployments call the scoring script's run(mini_batch) function, which receives a list of file paths and returns a pandas DataFrame or list of scored results.

MLflow no-code deployment

Deploying a registered model in MLflow format with a logged signature and environment, letting Azure ML generate the scoring script and container image automatically.

Model registry versioning

Each registration under the same model name creates an incremented version, so deployments and jobs pin an immutable version number or an assigned label such as production.

Deployment instance type

The VM SKU, such as Standard_DS3_v2, assigned to each online deployment instance; it determines the CPU and memory available to the model and scoring script.

Endpoint auth_mode setting

An endpoint's authentication mode: key uses primary and secondary keys for invocation, while aml_token requires an Entra ID bearer token obtained from the workspace.

Retrieval augmented generation

A pattern that chunks documents, embeds and indexes them, retrieves the most similar chunks per query, and injects them into the prompt to ground model answers.

Temperature parameter

A sampling setting from 0 to 2 controlling output randomness; values near 0 yield deterministic, repeatable answers, while higher values increase variation and creativity.

Fine-tuning data format

Supervised fine-tuning uses JSONL files where each line holds a messages array of system, user, and assistant turns, typically dozens to hundreds of high-quality examples.

Provisioned throughput units

Reserved model processing capacity purchased in PTU increments that guarantees consistent latency and throughput for a deployment regardless of demand from other customers.

Content filter severity levels

Azure OpenAI rates categories such as hate, sexual, violence, and self-harm as safe, low, medium, or high, blocking prompts or completions at or above the configured threshold.

Groundedness metric

An evaluation measure of how well a generated answer is supported by the retrieved or supplied source content, used to detect unsupported claims and hallucination.

Embedding vectors

Numeric vector representations of text produced by an embedding model; similarity search over these vectors powers semantic retrieval inside vector indexes such as Azure AI Search.

System message role

The leading instruction in a chat prompt that sets the assistant's persona, constraints, and expected output format and takes precedence over later user turns.