terraform-patterns
Terraform and OpenTofu infrastructure-as-code patterns for project layout, remote state, modules, variables, secrets, safe plan/apply workflows, and CI. Use when writing or reviewing .tf files, structuring modules or environments, configuring backends and state locking, or debugging plan drift, state, and provider errors.Terraform Patterns
Production-grade Terraform (and OpenTofu) patterns for writing infrastructure code that is reviewable, repeatable, and safe to apply.
When to Activate
- Writing or reviewing
.tfor.tfvarsfiles - Structuring a repo into modules and environments
- Configuring a remote backend, state locking, or state migration
- Passing variables and handling secrets in infrastructure code
- Setting up plan/apply automation in CI
- Debugging a failed plan, state drift, a lock error, or a provider version conflict
- Refactoring resources without destroying them (
moved,import)
When to Use
Same as When to Activate above. This alias satisfies repo skill-format conventions. Use this skill any time you are writing, reviewing, or debugging Terraform or OpenTofu configuration.
How It Works
This skill gives copy-pasteable patterns and a safe command workflow, organized by task:
- Project layout - Environments as thin roots that call reusable modules.
- Versions and providers - Pin Terraform and provider versions; commit the lock file.
- Remote state - Encrypted, locked, per-environment state.
- Variables and outputs - Typed, validated inputs and minimal outputs.
- Secrets - Keep values out of code, state exposure in mind.
- Resources -
for_eachovercount,lifecycle, and tagging. - Modules - Small, single-purpose, versioned.
- Refactoring -
movedandimportblocks instead of destroy/recreate. - Safe workflow and CI - Plan is saved, reviewed, then applied.
- Debugging and anti-patterns - Common errors and what to avoid.
Examples use AWS for concreteness. The structure applies to any provider.
Examples
| Task | Jump to |
|---|---|
| Layout for dev/staging/prod | Project Layout |
| Remote state with locking | Remote State |
| Validated variables | Variables and Outputs |
| Avoid secrets in state and code | Secrets |
| Rename a resource without destroying it | Refactoring Without Destroying |
| Plan/apply in CI | Safe Workflow and CI |
| Lock or drift errors | Debugging |
Project Layout
Keep environments as thin root modules. Put logic in reusable modules.
infra/
├── modules/
│ └── network/
│ ├── main.tf
│ ├── variables.tf
│ ├── outputs.tf
│ └── versions.tf
└── envs/
├── dev/
│ ├── main.tf
│ ├── backend.tf
│ ├── providers.tf
│ └── terraform.tfvars
└── prod/
├── main.tf
├── backend.tf
├── providers.tf
└── terraform.tfvars
Rules of thumb:
- One state per environment and per blast-radius boundary (network, data, apps). A single giant state makes every plan slow and every mistake large.
- Environments differ by inputs (
terraform.tfvars), not by copy-pasted resources. - Do not use workspaces as a substitute for separate production and non-production state with separate credentials.
Versions and Providers
Pin everything. Unpinned providers change behavior between runs.
# versions.tf
terraform {
required_version = ">= 1.10.0, < 2.0.0"
required_providers {
aws = {
source = "hashicorp/aws"
version = "~> 5.40"
}
}
}
# providers.tf
provider "aws" {
region = var.region
default_tags {
tags = {
Environment = var.environment
ManagedBy = "terraform"
Owner = var.owner
}
}
}
- Commit
.terraform.lock.hcl. It records provider checksums so CI and teammates install identical providers. - Do not commit the
.terraform/directory or any*.tfstatefile. - Upgrade deliberately:
terraform init -upgrade, review the lock file diff, then plan.
Remote State
Local state does not work for teams and is easy to lose. Use a remote backend with encryption and locking.
# backend.tf
terraform {
backend "s3" {
bucket = "example-org-tfstate-prod"
key = "network/terraform.tfstate"
region = "us-east-1"
encrypt = true
use_lockfile = true # S3-native locking (Terraform >= 1.10)
}
}
Older Terraform versions use a DynamoDB table for locking instead. Replace
use_lockfile = true with:
dynamodb_table = "example-org-tfstate-lock"
Backend rules:
- Backend blocks cannot reference variables. Pass values with
-backend-configor a partial config file when you need per-environment values. - Enable bucket versioning so a bad state write can be recovered.
- Block public access on the state bucket and restrict who can read it. State often contains sensitive values.
- Never use
-lock=falseto get past a lock error. See Debugging.
Variables and Outputs
Type and validate inputs so mistakes fail at plan time.
variable "environment" {
type = string
description = "Deployment environment."
validation {
condition = contains(["dev", "staging", "prod"], var.environment)
error_message = "environment must be one of: dev, staging, prod."
}
}
variable "instance_count" {
type = number
description = "Number of application instances."
default = 2
validation {
condition = var.instance_count >= 1 && var.instance_count <= 20
error_message = "instance_count must be between 1 and 20."
}
}
variable "allowed_cidrs" {
type = list(string)
description = "CIDR blocks allowed to reach the service."
default = []
}
output "vpc_id" {
description = "ID of the VPC."
value = aws_vpc.this.id
}
output "db_password" {
description = "Generated database password."
value = random_password.db.result
sensitive = true
}
- Every variable and output gets a
description. Modules are read far more often than written. - Give a
defaultonly when one is safe in every environment. Required inputs force the caller to decide. - Mark outputs that carry credentials
sensitive = true. Note this only hides them from CLI output; they are still stored in state.
Secrets
Terraform state stores resource attributes in plain text, including passwords and keys that you pass in or that the provider generates.
Do:
- Keep secrets out of
.tfand committed.tfvarsfiles. Supply them at runtime withTF_VAR_*environment variables or your CI secret store. - Reference secrets that already exist in a secret manager instead of creating the value in Terraform.
- Use ephemeral values and write-only attributes where your Terraform and provider versions support them, so the value never lands in state.
- Encrypt state at rest and restrict read access to it.
- Authenticate with short-lived credentials (OIDC from CI, assumed roles) rather than long-lived access keys.
# Read an existing secret instead of defining its value in code
data "aws_secretsmanager_secret_version" "db" {
secret_id = "prod/app/db-password"
}
resource "aws_db_instance" "app" {
identifier = "app-${var.environment}"
# ...
password = data.aws_secretsmanager_secret_version.db.secret_string
}
Reading a secret through a data source still places the value in state. Treat the state file as sensitive in every case.
Do not:
- Hardcode credentials in
providerblocks. - Commit
terraform.tfvarsfiles that contain secrets. Add them to.gitignore. - Print secrets with
terraform output -jsonin shared CI logs.
Resources
Prefer for_each over count
count identifies resources by index. Removing an item from the middle of a list shifts every later item down one index, so Terraform plans to modify or replace the resources after it (and destroy the last one) instead of just removing the one you meant to.
# Fragile: removing "b" shifts "c" into b's slot and destroys the last index
resource "aws_iam_user" "bad" {
count = length(var.users)
name = var.users[count.index]
}
# Stable: each resource is keyed by its name; removing "b" destroys only "b"
resource "aws_iam_user" "good" {
for_each = toset(var.users)
name = each.key
}
Use count only for a simple on/off toggle:
resource "aws_cloudwatch_log_group" "app" {
count = var.enable_logging ? 1 : 0
name = "/app/${var.environment}"
}
Lifecycle guards
resource "aws_s3_bucket" "data" {
bucket = "example-org-data-${var.environment}"
lifecycle {
prevent_destroy = true
}
}
resource "aws_launch_template" "app" {
# ...
lifecycle {
create_before_destroy = true
}
}
prevent_destroyon stateful resources (databases, buckets, encryption keys) turns an accidental destroy into a plan error.prevent_destroyonly protects a resource that is still in the configuration. If you delete the resource block, Terraform no longer sees the setting and will plan to destroy the real resource. To stop managing a resource without destroying it, use aremovedblock (Terraform >= 1.7):
removed {
from = aws_s3_bucket.data
lifecycle {
destroy = false
}
}
ignore_changeshides drift. Use it narrowly and add a comment saying why.
Secure defaults
resource "aws_s3_bucket_public_access_block" "data" {
bucket = aws_s3_bucket.data.id
block_public_acls = true
block_public_policy = true
ignore_public_acls = true
restrict_public_buckets = true
}
Avoid 0.0.0.0/0 ingress rules except for deliberately public endpoints, and keep IAM policies to the specific actions and resources needed.
Modules
A good module does one thing, exposes a small interface, and is versioned.
module "network" {
source = "../../modules/network"
environment = var.environment
cidr_block = "10.20.0.0/16"
}
# Registry or git modules: pin a version or ref
module "vpc" {
source = "terraform-aws-modules/vpc/aws"
version = "~> 5.0"
}
- Pin registry modules with
versionand git modules with?ref=<tag>. Never track a moving branch. - Do not configure
providerblocks inside child modules. Declarerequired_providersand let the root pass providers in. - Prefer composition: a module takes IDs as inputs rather than creating everything itself.
- Do not wrap a single resource in a module that only forwards its arguments.
Refactoring Without Destroying
Renaming a resource or moving it into a module looks like a destroy plus a create. Tell Terraform it is the same object.
# Resource renamed in code: keep the existing object
moved {
from = aws_instance.web
to = aws_instance.app
}
# Adopt an existing resource into state (Terraform >= 1.5)
import {
to = aws_s3_bucket.legacy
id = "example-org-legacy-bucket"
}
Always read the plan after a refactor. A correct moved block shows as "has moved" with no destroy.
Use terraform state commands only as a last resort, take a backup first, and never edit the state file by hand.
Safe Workflow and CI
Format and validate locally, then plan, review, and apply the same plan.
terraform fmt -recursive
terraform init
terraform validate
terraform plan -out=tfplan
terraform show tfplan # review before applying
# STOP: get explicit human approval of the reviewed plan before the next command
terraform apply tfplan # applies exactly what was reviewed
An AI agent must not run terraform apply or terraform destroy on its own. Show the plan summary, wait for the user to approve it, and only then apply the saved plan.
Why -out: a bare terraform apply re-plans, so what runs can differ from what you reviewed.
Read the plan for:
-/+(replace) and-(destroy) lines on stateful resources- Changes you did not make, which indicate drift or a provider upgrade
- Unexpected counts in the summary line
CI pipeline shape
# .github/workflows/terraform.yml (shape only)
name: terraform
on:
pull_request:
paths: ["infra/**"]
permissions:
id-token: write # OIDC to the cloud provider, no stored keys
contents: read
pull-requests: write
jobs:
plan:
runs-on: ubuntu-latest
steps:
# Pin third-party actions to full commit SHAs, not movable tags
- uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
- uses: hashicorp/setup-terraform@dfe3c3f87815947d99a8997f908cb6525fc44e9e # v4.0.1
with:
terraform_version: 1.10.5 # must satisfy required_version and use_lockfile (>= 1.10)
- run: terraform fmt -check -recursive
working-directory: infra/envs/dev
- run: terraform init -input=false
working-directory: infra/envs/dev
- run: terraform validate
working-directory: infra/envs/dev
- run: terraform plan -input=false -out=tfplan
working-directory: infra/envs/dev
- Run
planon pull requests and post it for review. Runapplyonly from the main branch, after approval, using the saved plan. - Use
-input=falsein automation so a missing variable fails instead of hanging. - Give the plan job read-only credentials where the provider allows it, and keep the apply role separate.
- Consider a policy or static-analysis step (for example
tflint,checkov, ortrivy config) before apply.
Debugging
| Symptom | Likely cause | What to do |
|---|---|---|
Error acquiring the state lock | Another run holds the lock, or a run crashed | Confirm no run is active, then terraform force-unlock <LOCK_ID>. Do not disable locking. |
| Plan shows changes you did not make | Drift, a provider upgrade, or a changed default | terraform plan -refresh-only to see drift. Check the lock file diff. |
Provider produced inconsistent result | Provider bug or version mismatch | Pin and upgrade the provider; check its issue tracker. |
Cycle error | Two resources depend on each other | Break the loop with an intermediate resource or remove the redundant reference. |
| Resource already exists | It was created outside Terraform | Add an import block, then plan. |
for_each value depends on unknown values | Keys come from an attribute known only after apply | Key by a static value, or apply in two steps with -target as a one-time exception. |
| Destroy and recreate on rename | Missing moved block | Add a moved block, re-plan. |
Useful commands:
terraform state list # what is tracked
terraform state show aws_instance.app # inspect one resource
terraform plan -refresh-only # detect drift without proposing changes
terraform providers # which providers and versions are in use
TF_LOG=DEBUG terraform plan 2> debug.log # verbose logs; may contain secrets
TF_LOGoutput can contain credentials. Do not paste it into issues or CI logs without scrubbing it.
Anti-Patterns
| Anti-pattern | Why it hurts | Instead |
|---|---|---|
terraform apply without a saved plan | Applies something other than what was reviewed | plan -out, review, apply tfplan |
-auto-approve on a human workflow | Removes the last safety check | Reserve it for pipelines that apply a reviewed plan |
| Local state for shared infrastructure | Lost state, no locking | Remote backend with locking |
| One huge state for everything | Slow plans, large blast radius | Split by environment and domain |
Secrets in .tf or committed .tfvars | Leaked credentials | Runtime variables or a secret manager |
| Unpinned providers or modules | Surprise changes between runs | Version constraints plus the lock file |
count over a list of named things | Index shift causes destroys | for_each with stable keys |
-target as routine practice | Skips dependencies, leaves state inconsistent | Fix the root cause; use only as a one-time exception |
| Editing state by hand | Corrupts state | moved, import, or terraform state with a backup |
| Disabling locking to bypass a stuck lock | Concurrent writes corrupt state | Verify and force-unlock the specific lock |
Checklist
Before opening a PR:
-
terraform fmt -recursiveandterraform validatepass - Terraform and provider versions are pinned;
.terraform.lock.hclis committed - State is remote, encrypted, and locked
- No secrets in code,
.tfvars, or outputs that are notsensitive - Variables have types, descriptions, and validation where a bad value is plausible
- Stateful resources have
prevent_destroyor an explicit reason they do not (and removals use aremovedblock withdestroy = false) - A human approved the saved plan before it was applied
- The plan was read: no unexpected destroys or replacements
- Refactors use
movedorimport, not destroy and recreate - CI applies the saved, reviewed plan
Related Skills
kubernetes-patterns- cluster workloads once the infrastructure existsdocker-patterns- container images that Terraform-managed infrastructure runsdeployment-patterns- release and rollout strategiessecurity-review- broader security checklistdatabase-migrations- schema changes, distinct from infrastructure changes