IaC (Infrastructure as Code)
Infrastructure as Code manages infrastructure through reviewable configuration or programs rather than undocumented manual changes. The competency is expressing intended resources, controlling state and dependencies, and applying changes safely so environments can be reproduced and differences can be inspected before they affect a running AI workload.
What it is
Infrastructure definitions describe compute, networking, storage, permissions and other resources that a provisioning tool creates or updates. Declarative tools reconcile desired configuration with tracked or discovered state; imperative approaches execute explicit provisioning operations. Both need identity, ordering and failure handling. Version control makes proposed changes reviewable but does not make their application harmless. Infrastructure as Code is the practice, while Terraform is one implementation. For AI systems, definitions may include expensive accelerators, model endpoints, data stores and the access relationships between them.
What the work involves
The practitioner models resource dependencies, separates environment-specific parameters and defines ownership boundaries so two tools do not manage the same object unpredictably. They review proposed changes, protect state and secrets, and apply through a controlled identity. They test representative environments and plan recovery for updates that cannot simply be undone. The result is reproducible infrastructure with a comprehensible change process, including a record of what was applied and a way to detect drift between intended and running resources.
Illustrative example
A team creates a staging inference environment from versioned definitions. The configuration includes storage, a serving service, its identity and narrowly scoped access. Before applying an update, the engineer notices that a renamed resource would be replaced rather than modified. They revise the migration plan to preserve required state and check the resulting environment by making an authorized and an unauthorized request.
Limits and common mistakes
Configuration can reproduce a mistake just as consistently as a correct architecture. State files and generated plans may contain sensitive values, and a rollback of source does not reverse every destructive resource change. Manual edits can introduce drift, while provider changes affect behavior. Inspect plans, state ownership, permissions and migrations. Infrastructure as Code reduces hidden setup; it does not replace understanding of the resources being provisioned.
Prerequisites
Related skills
- → is part of: MLOps
- ← is an instance of: Terraform
Sources and further reading
- Terraform introduction
Provides an official example of Infrastructure as Code, declarative configuration and provisioning workflows.
- Terraform state
Supports resource mapping, dependency metadata and state-management concerns.
Last updated: 2026-10-10