docs(P01): complete infrastructure-as-code phase

---ci---
project: atelier
phase: 1
milestone: v0.2
status: complete
requirements:
  covered: [ATELIER-36, ATELIER-37, ATELIER-38, ATELIER-39, ATELIER-40]
  partial: []
---/ci---
This commit is contained in:
Jon Chery
2026-08-05 02:04:11 +00:00
parent 40e61e6b7b
commit f755435ffe
6 changed files with 308 additions and 4 deletions
+4 -4
View File
@@ -1,10 +1,10 @@
{
"phase": 0,
"stage": "plan",
"phase": 1,
"stage": "execute",
"milestone": "v0.2",
"phase_role": "pre_execution",
"phase_role": "execution",
"project": "atelier",
"attempts": 0,
"updated_at": "2026-08-05T01:30:00Z",
"updated_at": "2026-08-05T01:40:00Z",
"milestone_complete": false
}
@@ -0,0 +1,55 @@
# Infrastructure as Code — First Principles
## 1. The Principles
### P1. Declarative Intent
Describe the desired state, not the steps to reach it. The tool
reconciles current → desired. Imperative scripts describe how;
declarative config describes what.
### P2. Idempotence
Applying the same configuration twice yields the same result. A
second `apply` with no changes is a no-op, not an error. Idempotence
is what makes `plan` trustworthy.
### P3. State is Truth
The state file is the authoritative record of what the tool believes
exists. Drift between state and reality is a bug to be reconciled,
not tolerated. Lose state, lose the ability to reason about
infrastructure.
### P4. Plan Before Apply
Preview every change before mutating real infrastructure. `plan` is
the contract review; `apply` is the signature. No `apply` without a
read `plan`. The plan is the rollback rehearsal.
### P5. Version Everything
Configuration, state, providers, and modules are all versioned and
reproducible. A commit pins a complete, rebuildable world. Pin
providers; pin module sources; never `latest`.
### P6. Modules Compose
Encapsulate repeatable patterns as versioned modules. Compose
modules; do not copy them. A module is the unit of reuse, review,
and versioning — the IaC expression of composition.
### P7. Least Privilege Providers
Provider credentials are scoped to the minimum needed for the
declared resources. No account-wide admin keys in CI. One credential
per environment, per boundary.
### P8. Remote State with Locking
State is stored remotely with locking. Local state is for a single
developer on a throwaway sandbox. Concurrent `apply` without a lock
is data corruption waiting to happen.
### P9. Drift is Recoverable
`plan` detects drift; `apply` reconciles it. Manual mutation of
managed infrastructure is an incident, not a shortcut. Drift is
expected; unreconciled drift is the bug.
### P10. Secrets Never in Code
Secrets come from providers, external secret stores, or environment
variables — never hardcoded in HCL, never committed to the repo,
never written to state in plaintext. State is a secret-bearing
artifact; treat it accordingly.
+64
View File
@@ -0,0 +1,64 @@
# Modules — Derived Rules
> Derives from `domains/infrastructure-as-code/first-principles.md`. P6 (Modules Compose) lives here. Referenced by `terraform.md` and `opentofu.md`.
## Why Modules (P6 Modules Compose)
- A module is the unit of reuse, review, and versioning in IaC. It encapsulates a repeatable pattern behind a typed interface.
- Composition — building large from small — is the IaC expression of core C6 Composability. Without modules, every stack is a one-off; with modules, a stack is an assembly of reviewed parts.
- A good module has one job (a VPC, a database, a load balancer), a small typed surface, and no hidden side effects.
## Module Structure (P1 Declarative Intent, C2 Clarity)
- The conventional layout: `main.tf` (resources), `variables.tf` (inputs), `outputs.tf` (outputs), `versions.tf` (provider/version pins). A `README.md` is required for any published module.
- Inputs are typed and validated: `variable "name" { type = string, description = "...", validation { ... } }`. The description is the contract.
- Outputs are the module's interface to consumers. Mark sensitive outputs `sensitive = true`. Document non-obvious outputs in the description.
- A module does not declare a provider configuration unless it owns the provider. Most modules declare only `required_providers` (the constraint) and let the consumer configure the provider.
## Versioning (P5 Version Everything)
- Modules are versioned. The registry expects SemVer tags (`v1.0.0`). A consumer pins to a version or a range (`~> 1.0`).
- A breaking change bumps the major. An additive change bumps the minor. A fix bumps the patch. No silent breaking changes within a minor.
- Tag the module repo; the tag IS the version. Never `source = "git::...?ref=main"` in prod — unversioned modules drift.
## Source Patterns (P5 Version Everything)
| Source | When | Risk |
|--------|------|------|
| Registry (`<ns>/<name>/<provider>`) | Public, versioned, signed | Verify the publisher; pin the version |
| Git (`git::https://...?ref=v1.0.0`) | Private modules across repos | Pin to a tag, not a branch |
| Local (`./modules/networking`) | Monorepo, single repo | Re-reviewed on every change; no independent version |
| Inline (no module) | Trivial one-off | Becomes a copy-paste anti-pattern at scale |
- Local modules in a monorepo are fine — they trade independent versioning for co-evolution. The boundary is the review unit: if the module and the consumer always change together, local is correct.
- Cross-repo modules must be versioned via git tags; unversioned cross-repo modules are the worst case (drift without a version to pin).
## The Module-vs-Copy Boundary (P6 Modules Compose)
- If a block is used more than once, it is a module. If it is used once and will never be reused, inline is acceptable.
- If two copies differ in one attribute, that is a module with a variable, not two copies. The variable is the difference; the shared body is the module.
- If you find yourself copy-pasting a block and editing it, stop. The edit is a variable. The copy is a module call.
- A module that has grown to do many jobs should be split. A module with 20 variables is two modules.
## Composition (P6 Modules Compose, C6 Composability)
- Compose by calling modules from a root configuration: `module "vpc" { source = "...", version = "..." }`. The root is the assembly; the modules are the parts.
- Outputs of one module feed inputs of another: `module "app" { vpc_id = module.vpc.vpc_id }`. This is the composition edge.
- Avoid hidden coupling: a module should not reach into another module's state. If two modules must share state, promote the shared concern to the root or a parent module.
## Reviewing Modules (P4 Plan Before Apply)
- A module is reviewed once, at its version. Consumers trust the version pin. A module change requires a new version and a review of the diff.
- When a module changes, every consumer that bumps the version gets the change. Treat a module version bump as a real change: review the module diff, run the consumer's `plan`.
- A module with a breaking change must not auto-bump in consumers. Pin consumers to the old major until they explicitly migrate.
## What Violates Module Discipline
| Violation | Principle |
|-----------|-----------|
| Copy-pasted block with a one-line difference | P6 Modules Compose |
| `source = "git::...?ref=main"` in prod | P5 Version Everything |
| Module with 20 variables | P6 Modules Compose (split it) |
| Silent breaking change within a minor | P5 Version Everything |
| Module reaching into another module's state | C6 Composability, P1 Declarative Intent |
| Unpublished module with no README | C2 Clarity |
@@ -0,0 +1,50 @@
# OpenTofu — Derived Rules
> Derives from `domains/infrastructure-as-code/first-principles.md`. OpenTofu is the open-source fork of Terraform; this doc covers fork-specific governance, license, and migration. The shared HCL/state/module model is documented in `terraform.md`. See also `state.md` and `modules.md`.
## Fork Lineage (P5 Version Everything)
- OpenTofu is a 2023 fork of Terraform, created when HashiCorp switched Terraform from MPL-2.0 to the Business Source License (BUSL), which is not open source.
- OpenTofu is stewarded by the Linux Foundation under a genuinely open-source license. The fork's reason for existing is license neutrality.
- Both tools implement the same HCL configuration language, the same provider protocol, and the same state model. Configuration written for one runs on the other at the fork point; divergence accrues slowly over time.
## When to Choose OpenTofu (P7 Least Privilege Providers, supply-chain)
- **License neutrality matters:** if your organization cannot accept BUSL's "competitive use" ambiguity, OpenTofu removes it.
- **Supply-chain provenance:** Linux Foundation stewardship means no single vendor can relicense the tool out from under you.
- **Community governance:** features and fixes are accepted on merit, not vendor strategy.
- **When NOT to switch:** if you depend on HCP Terraform (HashiCorp's managed platform), BUSL-licensed providers, or provider features that have diverged since the fork, stay on Terraform. The decision is supply-chain, not syntax.
## CLI Parity (P1 Declarative Intent)
- `tofu init`, `tofu plan`, `tofu apply`, `tofu destroy` mirror `terraform init/plan/apply/destroy`.
- The lock file (`.terraform.lock.hcl``.tofu.lock.hcl`) is committed; it makes `init` reproducible.
- Workspaces, state backends, and module sources behave as in Terraform — see `terraform.md` and `state.md`.
## Registry Parity (P6 Modules Compose)
- OpenTofu can consume the Terraform Registry and the OpenTofu Registry. Module version pinning works identically.
- Some providers have BUSL licenses that OpenTofu cannot ship; verify a provider's license before adopting it. An MPL or Apache provider is portable; a BUSL provider is not.
- See `modules.md` for module structure, which is unchanged from Terraform.
## Migration from Terraform (P5 Version Everything, P9 Drift is Recoverable)
- `terraform state pull > state.json``tofu state push state.json` carries state across. Validate with `tofu plan` after the push — the plan should be empty.
- Rename the binary in CI: replace `terraform` with `tofu` in scripts. The lock file may need regeneration.
- Migrate one workspace at a time. Do not big-bang a migration; rehearse on a non-prod workspace first (P4 Plan Before Apply applies to the migration itself).
- Pin the OpenTofu version in CI. A migration is a versioned, reviewed change, not a quiet swap.
## Governance and Community (cross-link `domains/security/supply-chain.md`)
- OpenTofu's governance model — impartial, community-driven, layered, modular, backwards-compatible — is itself a supply-chain principle. A tool you cannot trust to remain open is a tool you cannot build on.
- This is the OpenTofu angle on `security/supply-chain.md`: license is a supply-chain property, not a legal footnote.
## What Violates OpenTofu Discipline
| Violation | Principle |
|-----------|-----------|
| Assuming OpenTofu == latest Terraform (unverified parity) | P5 Version Everything |
| Migrating prod state without a non-prod rehearsal | P4 Plan Before Apply |
| Adopting a BUSL-licensed provider into OpenTofu CI | P7 Least Privilege Providers, supply-chain |
| Quiet swap of `terraform` for `tofu` without a versioned change | P5 Version Everything |
| Losing state during migration | P3 State is Truth |
+77
View File
@@ -0,0 +1,77 @@
# State — Derived Rules
> Derives from `domains/infrastructure-as-code/first-principles.md`. State is the cross-cutting IaC concern: P3 (State is Truth) and P8 (Remote State with Locking) live here. Referenced by `terraform.md` and `opentofu.md`.
## Why State Matters (P3 State is Truth)
- The state file is the tool's memory. It records every resource it has claimed, every attribute it has set, and every dependency it has inferred.
- Without state, `plan` cannot compute a diff — it would have nothing to diff against. Lose state, lose the ability to reason about infrastructure safely.
- State can contain plaintext secrets (any sensitive resource attribute). Treat state as a secret-bearing artifact: encrypt at rest, restrict access, never commit it.
## Remote State is Mandatory (P8 Remote State with Locking)
- Local state (`terraform.tfstate` on disk) is acceptable only for a single developer on a throwaway sandbox. Any shared or production environment uses a remote backend.
- A remote backend provides: durability (state survives workstation loss), shared access (team members and CI read the same state), and locking (concurrent `apply` is serialized).
- No locking = data corruption. Two `apply` runs against the same unlocked state race; the loser's changes are silently overwritten.
## Backend Comparison (P8, C4 Locality)
| Backend | Locking | Encryption | Best for | Notes |
|---------|---------|------------|----------|-------|
| S3 + DynamoDB | DynamoDB | SSE-KMS | AWS-hosted | The canonical AWS backend; DynamoDB provides the lock |
| GCS | Built-in | CMEK | GCP-hosted | Native locking via GCS object versioning |
| Azure Blob | Lease | Customer key | Azure-hosted | Lease-based locking |
| HTTP (remote) | Server-side | Server-side | Self-hosted / on-prem | Requires a backend server (e.g., `terraform-backend`) |
| Local | None | None | Single-dev sandbox | Never for shared or prod |
| Consul | KV lock | — | Consul shops | Locking via Consul sessions |
| Postgres | TX | DB encryption | DBA-owned infra | Row-level locking |
- Pick one backend per environment family. Mixing backends across environments fragments operational knowledge (C4 Locality).
- The backend config is part of the configuration, not a runtime secret. Credentials for the backend are runtime secrets.
## State Isolation per Environment (P4 Plan Before Apply, C4 Locality)
- One state per environment. Never share a single state file across dev, staging, and prod. A `plan` against a shared state crosses environment boundaries — a prod change could appear in a dev plan.
- Isolation patterns: separate workspaces, separate state keys in the same backend, or separate backends entirely. Stricter isolation = safer (separate backends for prod vs non-prod).
- Name state keys by environment and stack: `env:/prod/Networking`, not `prod` or `state`.
## Locking Discipline (P8 Remote State with Locking)
- `terraform force-unlock` is for a stuck lock after a crashed run, not for impatience. Verify the run is actually dead before forcing.
- A forced unlock without verifying the other run is dead causes the corruption the lock prevents.
- In CI, set a lock timeout so a wedged job fails rather than hanging.
## Sensitive Values in State (P10 Secrets Never in Code)
- Any `sensitive = true` attribute is hidden from plan output but stored in state in plaintext (unless the provider encrypts it).
- Backends with at-rest encryption (S3 SSE-KMS, GCS CMEK) protect state at rest. Access to the state file itself is the boundary.
- Never log, print, or commit state. Never pipe `terraform show` to a public channel.
## State Commands (P3 State is Truth)
- `terraform state list` — enumerate resources in state. First step of any state investigation.
- `terraform state show <addr>` — inspect one resource's recorded attributes.
- `terraform state mv` — rename a resource's address without destroying and recreating it. Use when refactoring module structure.
- `terraform state rm` — stop managing a resource without destroying it. Use when handing a resource to another configuration.
- `terraform state pull` / `push` — export and import state. Used in migrations (see `opentofu.md`).
- `terraform import` — bring an existing resource under management by recording its state. The resource must already exist; `import` does not create.
- All `state` subcommands except `list` and `show` mutate state. Treat them as changes: review the intent, run in CI where possible, and commit the resulting config change that justifies the state move.
## Drift and Reconciliation (P9 Drift is Recoverable)
- `terraform plan` reports drift: resources that exist in state but were changed out-of-band, or resources in state that no longer exist in the provider.
- `terraform apply` reconciles drift by bringing reality back to the declared state.
- Manual changes to managed resources are the cause of drift. Treat a drift report as an incident: find who made the manual change and why, then close the access path or the gap that allowed it.
- `terraform plan -refresh=false` skips drift detection. Use only when you know state is current and you want a fast plan; never use it to hide drift.
## What Violates State Discipline
| Violation | Principle |
|-----------|-----------|
| Committed `terraform.tfstate` | P3 State is Truth, P10 Secrets |
| Local state in prod | P8 Remote State with Locking |
| `force-unlock` without verifying the dead run | P8 Remote State with Locking |
| Shared state across environments | P4 Plan Before Apply, C4 Locality |
| Unnamed state keys (`env:/prod`) | C4 Locality |
| Manual change to a managed resource | P9 Drift is Recoverable |
| `state rm` to "fix" a stuck resource | P3 State is Truth |
@@ -0,0 +1,58 @@
# Terraform — Derived Rules
> Derives from `domains/infrastructure-as-code/first-principles.md`. Applies P1P10 to Terraform specifically. See also `opentofu.md` (the open-source fork), `state.md`, and `modules.md`.
## HCL Structure (P1 Declarative Intent)
- Resources are declared, not scripted. A resource block states what should exist; Terraform reconciles it.
- `resource "aws_s3_bucket" "logs" { ... }` — the type and name are the identity; the body is the desired state.
- Data sources read existing state without claiming ownership: `data "aws_caller_identity" "current" {}`.
- Variables are the input contract; outputs are the interface to consumers. Both are typed.
## Providers (P5 Version Everything, P7 Least Privilege)
- Pin the provider version: `required_providers { aws = { source = "hashicorp/aws", version = "~> 5.0" } }`.
- A provider block configures credentials and region. Credentials come from environment, files, or a secrets manager — never inline.
- One credential set per environment. Do not reuse a prod credential in a dev workspace.
## The Core Workflow (P4 Plan Before Apply)
- `terraform init` — resolve providers and modules. Reproducible from the lock file (`.terraform.lock.hcl`), which is committed.
- `terraform plan` — preview the diff. Read it. Every line. The plan is the contract review.
- `terraform apply` — execute the plan. Requires a reviewed plan in CI; in interactive use, requires typing `yes`.
- `terraform destroy` — tear down. Treat `destroy` as a first-class operation with its own plan review; prod destroys are a change event, not a keystroke.
## Workspaces (P4 Locality of Environments)
- Workspaces separate state for the same configuration across environments (dev, staging, prod).
- Do not use workspaces to separate unrelated stacks — use separate configurations. A workspace is an environment axis, not a project axis.
- State is isolated per workspace (see `state.md`).
## State Backends (P3 State is Truth, P8 Remote State with Locking)
- Remote state is mandatory for any shared or production environment. See `state.md` for backend selection and locking.
- Never commit `terraform.tfstate` to the repo. It is a secret-bearing artifact and a source of drift.
- `terraform state` subcommands inspect and manipulate state directly — use sparingly, only for recovery.
## Registry and Modules (P6 Modules Compose)
- The Terraform Registry hosts versioned, signed modules. Reference modules by version: `source = "terraform-aws-modules/vpc/aws"`, `version = "5.x"`.
- Compose modules rather than copy-pasting blocks. A module is reviewed once and reused many times.
- See `modules.md` for module structure, versioning, and the module-vs-copy boundary.
## Secrets (P10 Secrets Never in Code)
- Secrets via provider data sources (`aws_secretsmanager_secret_version`), environment variables, or a dedicated secrets provider. Never a literal string in a resource block.
- State may contain plaintext secrets if a resource attribute is sensitive. Mark attributes `sensitive = true` to keep them out of plan output; use a backend that encrypts state at rest (see `state.md`).
## What Violates Terraform Discipline
| Violation | Principle |
|-----------|-----------|
| Unpinned provider (`source` without `version`) | P5 Version Everything |
| `terraform apply` without a read `plan` | P4 Plan Before Apply |
| Local state in a shared environment | P8 Remote State with Locking |
| Hardcoded secret in HCL | P10 Secrets Never in Code |
| Copy-pasted resource blocks instead of a module | P6 Modules Compose |
| Manual change to a managed resource | P9 Drift is Recoverable |
| Admin credentials in CI | P7 Least Privilege Providers |