Databricks Mosaic AI Model Serving and Governance Training Course

Course overview

Teams moving generative AI and predictive models from notebooks into production need more than a working endpoint. They must control which models can be deployed, protect sensitive data in prompts and responses, manage access by business role, monitor cost and quality, and provide an audit trail when a model decision is challenged. Databricks Mosaic AI Model Serving provides the deployment layer, while Unity Catalog, MLflow and Databricks governance controls provide the operating model. This course addresses the gap between experimentation and a governed, supportable AI service.

Participants learn to package and register models, create and configure Mosaic AI Model Serving endpoints, and serve custom models, feature-based models and foundation-model workloads. The course covers endpoint permissions, secrets, authentication, autoscaling choices, traffic routing, inference logging, AI Gateway controls, and usage monitoring. Participants use MLflow to track model versions and evaluation evidence, Unity Catalog to govern model assets and access, and Databricks deployment practices to promote approved releases across environments.

Delivery combines instructor demonstrations with guided work in a Databricks workspace, production scenarios and design reviews. Labs use realistic cases such as a customer-support assistant and a risk-scoring API, requiring participants to configure endpoint access, capture inference records, apply guardrails and investigate service behaviour. Each participant leaves with a documented model-serving implementation pack: an endpoint configuration, model registration and promotion path, governance checklist, monitoring design and a deployment runbook suitable for adaptation in their own organisation.

The course is designed for practitioners responsible for putting AI services into operation and for technical leads who must establish repeatable controls without slowing responsible delivery.

Course objectives

By the end of this course, participants will be able to:

• Configure Mosaic AI Model Serving endpoints for registered custom models and foundation-model workloads
• Register, version and promote model candidates through MLflow and Unity Catalog
• Apply Unity Catalog privileges, service principals and secret scopes to secure model-serving access
• Implement AI Gateway rate limits, usage tracking and guardrail policies for governed inference
• Capture and analyse inference logs to investigate latency, failures, cost and response quality
• Design a production endpoint strategy covering scaling, authentication, traffic routing and rollback
• Create a model release checklist with approval evidence, ownership and operational controls
• Produce a deployment runbook and monitoring plan for a governed Mosaic AI service

Benefits of attending

Target competencies

Who should attend

• Machine Learning Engineers — who deploy trained models and need repeatable production serving patterns
• Data Engineers — who build governed Databricks data products that supply or consume model endpoints
• AI Engineers — who operationalise LLM applications with controlled inference, observability and access
• MLOps Engineers — who standardise model release, monitoring and rollback processes across teams
• Data Platform Administrators — who administer Unity Catalog permissions, identities and workspace controls
• Technical Product Managers — who must define accountable operating requirements for AI-enabled services

Requirements and prerequisites

Participants should be comfortable navigating a Databricks workspace and working with notebooks, SQL or Python. They need a working understanding of machine learning model lifecycle concepts, including training, validation, versioning and inference, plus basic familiarity with REST APIs and cloud identity or access-control concepts. Experience using MLflow or Unity Catalog is helpful but not essential; the course introduces the features used in the labs. Participants do not need to build neural networks, fine-tune an LLM, write advanced Python, or hold cloud-certification credentials. A laptop able to access a supplied Databricks environment is required for practical work.

Training methodology

Instructor-led modules establish the architecture and governance decisions behind each serving pattern, followed by guided configuration in a Databricks lab environment. Participants register models, create endpoints, configure Unity Catalog access, examine inference records and diagnose deliberately introduced operational issues. Short case discussions compare customer-facing LLM services with internal predictive-model APIs, including their different risk controls. Teams conduct a deployment design review using a governance checklist, then complete an individual application plan that maps course practices to one proposed model-serving workload in their organisation.

Course outline

Day 1: Mosaic AI serving architecture and production design

• Mosaic AI Model Serving architecture and endpoint lifecycle
• Custom model, feature-based model and foundation-model serving patterns
• Online inference request and response design
• Endpoint compute, scaling and capacity planning considerations
• Authentication methods for application-to-endpoint access
• Latency, availability and cost trade-offs for AI services
• Production readiness criteria for model-serving workloads

Workshop: Participants assess a notebook-based support-classification model and produce an endpoint architecture decision record with service requirements and risks.

Day 2: Model packaging, registration and controlled release

• MLflow experiment tracking for deployable model artefacts
• MLflow model signatures, input examples and dependency packaging
• Unity Catalog model registration and model version management
• Aliases, tags and metadata for approved model versions
• Validation evidence and model evaluation records
• Environment promotion from development to test and production
• Rollback planning and release approval checkpoints

Workshop: Participants package a scored model, register it in Unity Catalog, assign release metadata and create a version-promotion and rollback plan.

Day 3: Secure endpoint configuration and AI governance

• Unity Catalog privileges for models, endpoints and governed data
• Service principals, personal tokens and application identity patterns
• Secret management for external model-provider credentials
• Endpoint access controls and least-privilege design
• AI Gateway usage tracking and rate-limiting controls
• Guardrail approaches for sensitive prompts and unsafe responses
• Audit requirements for model ownership, approvals and access changes

Workshop: Participants configure role-based access for a shared endpoint and produce a governance control matrix for users, applications and administrators.

Day 4: Inference observability, quality and operational response

• Inference tables and request-response logging design
• Latency, error-rate and throughput measures for serving endpoints
• Token usage and cost attribution for foundation-model traffic
• Tracing and evaluation evidence for LLM application responses
• Data privacy decisions for prompts, outputs and retained logs
• Failure analysis for authentication, dependency and capacity issues
• Incident triage, escalation and service-level operating procedures

Workshop: Participants investigate a simulated endpoint incident using inference records and produce an incident report with corrective actions and monitoring thresholds.

Day 5: Deployment automation and governed service operating model

• Databricks Asset Bundles for repeatable deployment definitions
• Parameterising endpoint configuration across environments
• Continuous integration checks for model and endpoint changes
• Separation of duties in AI release workflows
• Traffic migration, canary release and rollback decision criteria
• Model-serving runbook structure and ownership model
• Post-deployment review and control-improvement cycle

Workshop: Participants assemble and present a governed Mosaic AI deployment pack containing endpoint configuration, release checklist, monitoring plan and operational runbook.

Tools & standards covered

Databricks Mosaic AI Model Serving, Unity Catalog, MLflow, Databricks Asset Bundles

A typical training day

Live online deliveries follow the same structure in the East Africa Time zone, with shorter screen blocks and longer breaks.

What the fee includes

How you can take this course

Certification

Participants who complete the full five days receive the Skillset Development Certificate of Completion, stating the course title, course code, dates and delivery format — suitable for professional-development records and employer reimbursement.

Frequently asked questions