Azure DP-100 cheat sheet

The Azure ML workspace is the top-level resource; it groups compute, datastores, data assets, environments, components, models, endpoints, and experiments, and links to Key Vault, Storage, Container Registry, and Application Insights.

Compute selection rule: compute instance = single-user dev VM; compute cluster = autoscaling multi-node for training and sweeps; serverless compute = per-job managed allocation with no cluster to maintain; managed online endpoint = real-time inference; batch endpoint = scheduled or on-demand batch scoring.

Compute cluster parameters that control cost: min_nodes (0 enables scale-to-zero), max_nodes, idle_time_before_scale_down (default 120 seconds), vm_size, and tier (dedicated vs low-priority).

Datastore credential options: account key, SAS token, or workspace managed identity; the credential-less option stores no secret in the workspace and is required for private, key-free designs.

Data asset types: uri_file and uri_folder point at raw storage paths; mltable attaches a schema and supports tabular/parquet reads without custom parsing code; all asset versions are immutable.

An environment bundles a base Docker image, conda.yaml or pip requirements, and environment variables; curated azureml: environments are prebuilt, and registries can share environments across workspaces.

Identity decision rule: the workspace system-assigned managed identity is used by default for Storage and Key Vault access; a user-assigned identity is used when permissions must persist beyond the workspace lifecycle or be shared across resources.

Network isolation requires private endpoints for the workspace, storage account, Key Vault, and container registry plus a managed VNet (isolation_mode) or a bring-your-own VNet with delegated subnets and private DNS zones.

A command job is defined in YAML with type: command plus command, environment, compute, and code fields, then submitted with ml_client.jobs.create_or_update(Job(...)) or az ml job create.

MLflow tracking calls: mlflow.log_param, mlflow.log_metric, and mlflow.log_artifact write to the workspace-linked tracking server; mlflow.autolog() records parameters, metrics, and models automatically for scikit-learn, PyTorch, and XGBoost.

A sweep job uses type: sweep with search_space, sampling_algorithm (random, grid, or bayesian), objective.primary_metric plus objective.goal (maximize or minimize), and limits.max_total_trials.

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

Early-termination policies: bandit (slack_factor and evaluation_interval on the best-performing trial), median_stopping (compares against the median of running averages), and truncation_selection (cancel_percentage).

AutoML configuration keys: task, primary_metric, training_data, either validation_data or n_cross_validations, limits.timeout_minutes, max_concurrent_trials, and featurization mode (auto, off, or custom).

Data drift monitoring compares a baseline dataset to a target dataset over time windows using metrics such as population stability index and Jensen-Shannon distance, with alerts triggered on threshold breaches.

The responsible AI dashboard combines error analysis, Fairlearn-based fairness assessment, SHAP-based interpretability, and counterfactual what-if analysis for a registered model.

A scoring script for a managed online deployment must define init() (load model and dependencies once) and run(raw_data) returning a JSON-serializable value such as a dict or list.

Online deployment YAML keys: endpoint_name, model (registered asset path), environment, code_configuration.scoring_script, instance_type, instance_count, and scale_settings.

Invocation: ml_client.online_endpoints.invoke(endpoint_name=..., request_file=..., deployment_name=...) posts JSON to the endpoint's scoring URI and returns the model response.

Traffic control: the endpoint's traffic map assigns percentages to deployments and must total 100; blue-green rollout creates a new deployment, validates it with a small traffic share, then shifts to 100%.

Endpoint authentication: auth_mode key uses primary/secondary keys for calls, while auth_mode aml_token requires an Entra ID bearer token scoped to the workspace.

Autoscaling for managed online deployments uses scale type default with target_utilization_percentage, or scale type manual with a fixed instance_count.

Batch deployments run a scoring script's run(mini_batch) function, which receives a list of file paths and returns a pandas DataFrame or list written to the configured output location.

MLflow no-code deployment works when the registered model has an MLflow format, a logged signature, and a Python environment; Azure ML then generates the scoring script and image automatically.

Registering a model under an existing name increments its version; deployments and jobs pin an exact immutable version number or an assigned label such as latest or production.

Supervised fine-tuning data is JSONL with one object per line containing a messages array of system, user, and assistant turns; dozens of high-quality examples are the practical minimum, and hundreds typically improve results.

Retrieval augmented generation pipeline: chunk source documents, embed chunks with an embedding model, index vectors in Azure AI Search, retrieve top-k matches per query, and inject them into the prompt with instructions to cite sources.

Sampling parameters: temperature (0-2) controls randomness and should be near 0 for extraction or classification; top_p performs nucleus sampling and is usually adjusted instead of temperature, not both.

Token budgeting: prompt tokens plus max_tokens for the completion must fit inside the model's context window, so long retrieved context must be truncated or summarized before the call.

Provisioned throughput units (PTU) reserve model capacity for predictable latency and throughput, while standard and global-standard deployments bill per token on shared capacity.

Content filtering assigns severity levels safe, low, medium, and high across hate, sexual, violence, and self-harm categories, and blocks prompts or completions at or above the configured threshold.

Evaluation metrics for generated answers include groundedness (supported by retrieved context), relevance, coherence, and fluency; scores come from a judge model over a labeled test dataset.

Embedding models convert text into fixed-length vectors; semantic search ranks results by cosine similarity between the query vector and indexed document vectors.