AI and ML

AI and ML platform guide

guideazure2 min read4 days old

Use this page to find guidance for machine learning and generative AI workloads on Azure Databricks. Product names and preview status can change. Check the current Azure documentation and Azure release notes before you select a feature.

Platform map#

Need Start with Design concern
Experiments and model lifecycle MLflow Reproducible code, data, parameters, metrics, and promotion evidence
Governed models and functions Unity Catalog Ownership, environment boundaries, lineage, and least privilege
Online inference Model Serving Authentication, latency, scaling, observability, and cost limits
Retrieval for an AI application Vector Search Document permissions, refresh strategy, evaluation, and citation quality
Production agent or RAG system MLflow evaluation plus serving Quality gates, trace review, unsafe-output handling, and rollback

Do not start with the newest feature. Start with the workload's data boundary, quality target, latency, expected traffic, and failure mode. Those requirements determine whether a notebook, batch job, SQL function, model endpoint, or application is the right delivery surface.

Production baseline#

  1. Govern inputs and outputs. Put source data, features, models, functions, and evaluation datasets under explicit Unity Catalog ownership. Preserve the caller's permissions when a retrieval path can expose governed data.
  2. Separate experimentation from promotion. Development can iterate quickly; production needs a versioned artifact, an evaluation record, an owner, and a rollback target.
  3. Evaluate the actual task. Use representative business examples and measure the errors that matter. Generic model scores do not prove that an extraction, recommendation, or agent workflow meets the release requirements.
  4. Trace and monitor. Capture inputs, outputs, latency, failures, model or prompt version, and retrieval context without logging secrets or unnecessary personal data.
  5. Limit cost and the effect of a failure. Set endpoint ownership, budgets, rate limits, timeouts, and a fallback path before opening the workload to broad use.
  6. Treat preview features as replaceable. Isolate them behind a small interface and record the feature status checked during design review.

Review questions#

  • What data may the workload read, and can the answer reveal data the caller could not query?
  • What test set blocks a bad model, prompt, retrieval change, or agent tool from promotion?
  • Which identity invokes the workload, and which identity reaches downstream data or services?
  • Where are traces retained, who can read them, and how are sensitive values removed?
  • What happens when the model is slow, unavailable, over budget, or confidently wrong?
  • Can the team reproduce and roll back the exact production version?

Official sources#

Referenced by