terraform（Terraform 基础设施）

Terraform & OpenTofu: Production Infrastructure-as-Code

Write, review, and architect Terraform/OpenTofu infrastructure - from individual resources to multi-account, PCI-compliant platform architectures. The goal is reproducible, drift-free, auditable infrastructure that passes both peer review and QSA assessment.

Target versions (April 2026): Terraform 1.14.9 (IBM/HashiCorp, BSL; 1.15.0-rc2 in progress), OpenTofu 1.11.6 (Linux Foundation, MPL). Helm provider v3.1+, K8s provider v3.0+, AWS provider v6.x, Azure v4.x, GCP v7.x.

This skill covers four domains depending on context:

• HCL - resource configs, variables, outputs, data sources, expressions, lifecycle rules
• Modules - structure, versioning, testing, registry patterns, reusable components
• Operations - state management, backends, workspaces, import, migration, CI/CD
• Compliance - PCI-DSS 4.0 controls, policy-as-code, audit trails, drift detection, CDE isolation

Terraform vs OpenTofu (2026)

IBM acquired HashiCorp for $6.4B (closed Feb 2025). Terraform stays BSL 1.1. The codebases have meaningfully diverged.

Choose Terraform if: already on HCP Terraform/TFE, need Stacks for multi-component orchestration, want vendor support.

Choose OpenTofu if: need client-side state encryption (Terraform never shipped this), BSL is a legal concern, want enabled meta-argument on resources, want OCI registry for providers/modules, need Linux Foundation governance.

Both share the provider plugin protocol - most providers work on both. For now.

CDKTF is dead. Deprecated Dec 2025, archived. Migrate to HCL or AWS CDK.

When to use

• Writing or reviewing Terraform/OpenTofu configurations
• Designing module architecture or registry patterns
• Planning state management, backend strategy, or migration
• Setting up CI/CD pipelines for IaC (plan/apply workflows)
• Implementing policy-as-code gates (Checkov, OPA, Sentinel)
• PCI-DSS 4.0 compliance for infrastructure provisioning
• Multi-account/multi-cloud architecture with blast radius controls
• Reviewing AI-generated Terraform for security and correctness

When NOT to use

• Kubernetes manifests or Helm charts (use kubernetes)
• Ansible playbooks or configuration management (use ansible)
• Docker/container optimization (use docker)
• CI/CD pipeline design (use ci-cd)
• Database engine configuration, schema design, or migrations (use databases)
• Security auditing application code (use security-audit)

AI Self-Check

AI tools consistently produce the same Terraform mistakes. Before returning any generated HCL, verify against this list:

• No hardcoded values - regions, AMI IDs, CIDR blocks, account IDs must be variables
• No overly permissive IAM - no "Action": "*" or "Resource": "*" unless explicitly requested
• No 0.0.0.0/0 ingress on security groups (except port 443 for public ALBs, justified)
• S3 buckets: aws_s3_bucket_public_access_block with all four settings true (unless public access is explicitly required and justified), plus SSE-KMS encryption (aws_s3_bucket_server_side_encryption_configuration), versioning enabled, access logging (aws_s3_bucket_logging), and no overly permissive bucket policy (review aws_s3_bucket_policy for broad Principal: "*" grants)
• Provider versions pinned in required_providers with ~> constraints
• Backend config present (not local) with encryption and locking
• lifecycle blocks where needed (create_before_destroy, prevent_destroy on stateful resources)
• sensitive = true on variables/outputs containing secrets
• Tags on every taggable resource (at minimum: Name, Environment, Owner, pci_scope if applicable)
• No deprecated resource arguments (check provider changelog - AI trains on old syntax)
• No provisioner blocks - use Ansible or user_data instead
• State file does NOT contain plaintext secrets (use ephemeral resources on TF 1.10+ or data sources for runtime secret lookup)
• terraform fmt and terraform validate pass

AI should never own terraform apply. In March 2026, an AI-assisted Terraform workflow deleted production infrastructure through escalating cleanup logic. Plan output is reviewed by a human. Always.

Workflow

Step 1: Determine the domain

Based on the request:

• "Create a VPC/RDS/EC2/resource" -> HCL
• "Create a reusable module" -> Modules
• "Set up state backend" / "migrate state" -> Operations
• "Make this PCI compliant" / "policy gates" -> Compliance
• "Review this Terraform" -> Apply production checklist + critical rules + AI self-check
• "Review S3 buckets" -> S3 hardening review (see below) + AI self-check

Step 2: Gather requirements

Before writing HCL, determine:

• Cloud provider(s) and account/project structure
• Resource type and its dependencies
• Environment (dev/staging/prod) and promotion strategy
• State backend and locking mechanism
• Compliance scope: PCI CDE? Regulated? What tags/policies apply?
• Existing modules: reuse before creating new ones
• Secrets: how are they injected? (Vault, SSM, Secrets Manager - never tfvars)

Step 3: Build

Follow the domain-specific section below. Always terraform fmt + terraform validate + run Checkov before finishing.

Step 4: Validate

terraform fmt -check -recursive              # Format check
terraform validate                            # Syntax + provider validation
tflint --recursive                           # Provider-specific linting
checkov -d . --framework terraform           # Security/compliance scan
terraform plan -out=plan.tfplan              # Review the plan
terraform show -json plan.tfplan | conftest test -  # Policy-as-code gate (OPA)

HCL Patterns

Resource structure

resource "aws_instance" "web" {
  ami           = var.ami_id
  instance_type = var.instance_type
  subnet_id     = var.private_subnet_id

  root_block_device {
    encrypted   = true
    kms_key_id  = var.kms_key_arn
    volume_size = 20
  }

  metadata_options {
    http_endpoint = "enabled"
    http_tokens   = "required"  # IMDSv2 - enforce this always
  }

  tags = merge(var.common_tags, {
    Name = "${var.project}-web-${var.environment}"
  })

  lifecycle {
    create_before_destroy = true
  }
}

Key patterns

Variables: type them. Default non-sensitive ones. Mark secrets sensitive. Use validation blocks for constraints.

variable "environment" {
  type        = string
  description = "Deployment environment"
  validation {
    condition     = contains(["dev", "staging", "prod"], var.environment)
    error_message = "Must be dev, staging, or prod."
  }
}

variable "db_password" {
  type      = string
  sensitive = true  # prevents logging in plan output
}

Locals: extract repeated expressions. Name descriptively.

locals {
  name_prefix = "${var.project}-${var.environment}"
  is_prod     = var.environment == "prod"
  common_tags = {
    Project     = var.project
    Environment = var.environment
    ManagedBy   = "terraform"
    pci_scope   = var.pci_scope
  }
}

Data sources: for runtime lookups. Never hardcode AMI IDs, AZ lists, or account IDs.

data "aws_caller_identity" "current" {}
data "aws_availability_zones" "available" { state = "available" }
data "aws_ami" "al2023" {
  most_recent = true
  owners      = ["amazon"]
  filter {
    name   = "name"
    values = ["al2023-ami-*-x86_64"]
  }
}

Ephemeral resources (TF 1.10+ / OT 1.11+): secrets that never persist in state. Ephemeral values can only flow into write_only arguments, provider configs, provisioners, or other ephemeral contexts - not into regular resource arguments.

ephemeral "aws_secretsmanager_secret_version" "db_password" {
  secret_id = "prod/db/master-password"
}

resource "aws_db_instance" "main" {
  password = ephemeral.aws_secretsmanager_secret_version.db_password.secret_string  # must be write_only in provider
}

Lifecycle rules: use deliberately, not defensively.

• create_before_destroy - for zero-downtime replacements (LBs, ASGs, DNS)
• prevent_destroy - for stateful resources (databases, S3 buckets with data)
• ignore_changes - for attributes managed outside Terraform (ASG desired_count managed by HPA)
• replace_triggered_by - force recreation when a dependency changes

Import blocks (TF 1.5+): declarative imports, no state surgery.

import {
  to = aws_s3_bucket.existing
  id = "my-existing-bucket"
}

Moved blocks (TF 1.8+): declarative intra-state refactoring. Rename resources or move into/out of modules within the same state file. Reviewed in PRs, applied automatically on terraform apply. Does NOT work across state files.

moved {
  from = aws_instance.web
  to   = module.compute.aws_instance.web
}

Cross-state resource move (state surgery - when moved blocks can't help):

# 1. Back up source state, then remove the resource
terraform state pull > backup.tfstate         # safety backup only
terraform state rm aws_instance.web           # removes from source backend directly

# 2. In the destination workspace, import the resource
terraform import aws_instance.web i-0abc1234def56789
# Then add the matching resource block in HCL to avoid drift

Verify both states with terraform plan before and after. state rm writes directly to the backend - do not state push the backup afterward (that would undo the removal). Ensure no other runs hold the state lock before starting (check terraform force-unlock only as a last resort with a known-stale lock ID) and block concurrent apply in CI for both source and destination during the migration.

What NOT to write

• provisioner "local-exec" or provisioner "remote-exec" - use Ansible
• depends_on when Terraform already infers the dependency from attribute references
• count for conditional resources when for_each with a set is clearer (OpenTofu: use enabled)
• String interpolation for simple references: "${var.name}" -> var.name
• terraform.tfvars committed to Git with real values
• terraform.workspace for environment separation (use separate state files or workspaces with distinct backends)
• Inline provisioner blocks of any kind

S3 bucket review checklist

When reviewing or writing S3 bucket configurations, verify every bucket has all six companion resources. AI-generated HCL routinely omits several of these.

Also verify the account-level safety net: aws_s3_account_public_access_block with all four settings true. This catches any bucket that accidentally ships without its own block.

For PCI CDE buckets, add aws_s3_bucket_object_lock_configuration with COMPLIANCE mode retention for immutable audit storage (Req 10.5).

See references/compliance.md for full S3 hardening HCL examples including account-level blocks and object lock.

Modules

Read references/module-patterns.md for detailed module structure, testing patterns, and registry strategies.

Structure

modules/<provider>/<resource-type>/
  main.tf           # Resources
  variables.tf      # Inputs
  outputs.tf        # Outputs
  versions.tf       # Required providers + terraform version
  README.md         # Usage examples
  examples/         # Working example configs
  tests/            # .tftest.hcl files

Versioning

• Production: pin exact versions (= 2.1.3) or use dependency lock file
• Dev/staging: allow minor updates (~> 2.1)
• Every module gets semantic versioning and a CHANGELOG
• Provider versions: pin with ~> in required_providers. The .terraform.lock.hcl file pins exact hashes - commit it.

Testing (2026 standard)

• terraform test (native, GA): HCL-based unit tests for every module. Fast, runs in CI on every PR.
• Terratest (Go): integration tests that spin up real infrastructure. Run nightly or pre-release.
• Both complement each other. terraform test for fast validation, Terratest for real-world proof.

Anti-patterns

• Modules wrapping a single resource with no added logic (just use the resource directly)
• Modules with more variables than the resource they wrap has arguments
• module "vpc" that just passes through all variables to aws_vpc
• Not pinning module versions in production
• Using Git refs for module sources in production (use a registry or exact tags)

Operations

Read references/state-and-security.md for state backends, locking, encryption, OIDC federation, CI/CD pipeline patterns, and state surgery (cross-state resource migration).

State management

See references/state-and-security.md for full backend config examples, OIDC federation patterns, CI/CD pipeline flows, and cross-state migration workflows.

S3 + native locking (TF 1.10+): DynamoDB-based locking is deprecated. Use use_lockfile = true. Encrypt with KMS. Enable versioning and CloudTrail data events on the bucket.

OpenTofu: add client-side state encryption on top (AES-GCM, AWS KMS, GCP KMS, or OpenBao) - encrypts before upload, even a compromised backend can't read state.

State file splitting (blast radius)

Split by risk and ownership:

states/
  network/cde/         # CDE VPC - separate IAM role, separate approval
  network/non-cde/     # Everything else
  compute/cde/         # Payment processing
  compute/non-cde/     # App tier
  data/cde/            # RDS with cardholder data
  iam/                 # IAM is high-risk - own state, own approval
  monitoring/          # CloudTrail, GuardDuty, Config

CDE state files get their own backend, IAM role, and approval workflow. A terraform apply on non-CDE infra must never touch CDE resources.

CI/CD credentials: OIDC federation

No static credentials in CI. Use OIDC federation (GitHub Actions, GitLab CI):

• CI generates a signed JWT per pipeline run
• Cloud provider validates JWT against CI platform's OIDC endpoint
• Short-lived credentials issued, scoped to that execution
• Separate roles: read-only for plan (any branch), write for apply (main only)
• Lock subject claims to specific repos AND branches

Supply chain integrity

The Terraform ecosystem has real supply chain risks (March 2026):

• Pin GitHub Actions to commit SHAs - tj-actions/changed-files was compromised March 2025 via upstream reviewdog/action-setup (CVE-2025-30154) (~12 hours of credential theft). Same pattern as the Trivy compromise a year later.
• Module supply chain is weak - modules have no hash verification (unlike the provider lock file). Typosquatting on the public registry is a demonstrated attack vector (NDC Oslo 2025).
• Terrascan: dead. Archived Nov 2025. Migrate to Checkov or Trivy.
• tfsec: merged into Trivy. Still works standalone but no new development.
• Trivy IaC scanning: pin to a verified version in CI. Check release notes before updating - supply chain attacks on CI tools are real. Pin to SHA digest, not mutable tag.
• CDKTF: dead. Deprecated Dec 2025, archived. Migrate to HCL.

Architecture

Multi-account strategy

Organization root
+-- Security OU
|   +-- Log Archive account (CloudTrail, Config, audit logs)
|   +-- Security Tooling account (GuardDuty, Security Hub)
+-- Infrastructure OU
|   +-- Shared Services account (Transit Gateway, DNS, CI/CD)
+-- Workloads OU
|   +-- Dev account
|   +-- Staging account
|   +-- Production account
+-- CDE OU (PCI)
    +-- CDE Production account (payment processing - isolated)
    +-- CDE Staging account

Terraform manages cross-account via provider aliases with assume_role:

provider "aws" {
  alias  = "cde"
  region = var.region
  assume_role {
    role_arn = "arn:aws:iam::CDE_ACCOUNT:role/TerraformDeployRole"
  }
}

Security scanning stack

Recommended CI pipeline: terraform fmt -> terraform validate -> tflint -> checkov -> terraform plan -> conftest test (OPA) -> human review -> terraform apply

Compliance

Read references/compliance.md for the full PCI-DSS 4.0 requirements mapping, drift detection strategy, audit trail architecture, and OIDC patterns.

Quick reference: PCI-DSS 4.0 and IaC

PCI DSS 4.0 explicitly puts IaC repos in scope (Req 6). Your Terraform repo needs the same controls as any CDE system - access controls, audit logging, change management.

Critical requirements (most commonly cited in QSA findings):

• Req 6 + 8.6.2: IaC repo in scope - PR reviews, policy-as-code gates, no hardcoded secrets (use ephemeral resources TF 1.10+ or Vault/SSM)
• Req 10: Audit trail - archived plan/apply JSON + CloudTrail + immutable S3 with object lock (COMPLIANCE mode)
• Req 11.5: Drift detection satisfies FIM - schedule terraform plan runs and alert on unexpected changes

See references/compliance.md for full Req 1/3/7 mapping, drift detection strategy, and audit trail architecture.

State file security: state contains secrets (even with sensitive). Encrypt at rest (S3 SSE-KMS), restrict access (IAM policy), enable versioning, log all access (CloudTrail data events), retain 1+ year.

QSA expectations (2026): operational proof, not policy intent. Git history showing reviewed PRs, archived plan outputs, policy scan results per deployment, drift reports proving continuous compliance.

Production Checklist

Read references/production-checklist.md for the full pre-deploy checklist covering HCL quality, module standards, operations, PCI-DSS 4.0, and PCI MPoC compliance.

Reference Files

• references/module-patterns.md - module design and testing patterns
• references/state-and-security.md - state backend, locking, encryption, OIDC patterns, and state surgery (cross-state migration)
• references/compliance.md - compliance and audit-oriented Terraform guidance
• references/production-checklist.md - pre-deploy verification checklist (HCL, modules, operations, PCI-DSS, MPoC)

Related Skills

• ansible - for day-2 configuration of provisioned resources. Terraform provisions the VM; Ansible configures what runs on it. No provisioner blocks - use Ansible instead.
• kubernetes - K8s manifests and Helm charts. Terraform provisions the cluster; kubernetes configures what runs on it.
• databases - engine tuning and operations. Terraform provisions managed databases; databases skill tunes the engine.
• ci-cd - pipeline design that runs terraform plan/apply. Terraform covers HCL; ci-cd covers the pipeline stages.
• docker - container image patterns. Terraform provisions container infrastructure but Dockerfile design belongs in docker.

Rules

These are non-negotiable. Violating any of these is a bug.

1. terraform fmt + terraform validate on every change. Non-negotiable.
2. Pin provider versions. required_providers with ~> constraints. Commit the lock file.
3. Never commit secrets. Not in .tf, not in .tfvars, not in state. Use ephemeral resources, Vault, or SSM.
4. No provisioner blocks. Use Ansible or user_data.
5. **No "Action": "*" in IAM policies.** Least-privilege only.
6. Encrypt everything. Storage, transit, state backend. No exceptions.
7. State backend with locking and encryption. Never local state in production.
8. Separate CDE state files. Own backend, own IAM role, own approval workflow.
9. OIDC federation for CI/CD. No static cloud credentials.
10. Pin CI actions to commit SHAs. Mutable tags are compromised supply chain vectors (tj-actions March 2025, Trivy March 2026).
11. terraform plan before every apply. Archive the plan output.
12. AI never owns terraform apply. Plan output is reviewed by a human. Always.
13. Run the AI self-check. Every generated HCL gets verified against the checklist above before returning.