atmos-design-patterns（大气设计模式）

Atmos Design Patterns

Design patterns are proven solutions for structuring infrastructure configuration in Atmos. They address organizational complexity by providing reusable approaches for multi-account, multi-region, enterprise-grade environments.

Pattern Progression

Most teams follow this growth path:

Inline Configuration (learning/prototyping)
    |
Basic Stack Organization (dev/staging/prod)
    |
Multi-Region Configuration (add regions)
    |
Organizational Hierarchy (add teams/accounts/OUs)

Start with the simplest pattern that meets your needs. You do not need to start with the most complex pattern -- start simple and evolve.

Stack Organization Patterns

Basic Stack Organization

One file per environment. Simplest setup for single-region, single-account-per-stage deployments.

stacks/
  catalog/
    vpc/
      defaults.yaml      # Shared component defaults
  deploy/
    dev.yaml             # Imports catalog, sets stage: dev
    staging.yaml
    prod.yaml

Each environment file imports shared defaults and adds environment-specific overrides:

# stacks/deploy/dev.yaml
import:
  - catalog/vpc/defaults
vars:
  stage: dev
components:
  terraform:
    vpc:
      vars:
        nat_gateway_enabled: false

Deploy with: atmos terraform apply vpc -s dev

Multi-Region Configuration

Extends basic pattern to deploy across multiple AWS regions. Each region gets its own stack file with region-specific settings (CIDR blocks, availability zones).

stacks/deploy/dev/
  us-east-2.yaml          # region: us-east-2, environment: ue2
  us-west-2.yaml          # region: us-west-2, environment: uw2

Use name_template: "{{.vars.environment}}-{{.vars.stage}}" in atmos.yaml to generate stack names like ue2-dev.

Organizational Hierarchy Configuration

Enterprise pattern for multiple organizations, OUs/tenants, and accounts. Uses _defaults.yaml files at each hierarchy level to create inheritance chains.

stacks/orgs/acme/
  _defaults.yaml                    # namespace: acme
  plat/
    _defaults.yaml                  # tenant: plat (imports org defaults)
    dev/
      _defaults.yaml                # stage: dev (imports tenant defaults)
      network.yaml                  # layer: network (imports stage defaults + catalog)
      data.yaml                     # layer: data
    prod/
      _defaults.yaml
      network.yaml
      data.yaml
      compute.yaml

Import chain: network.yaml -> prod/_defaults.yaml -> plat/_defaults.yaml -> acme/_defaults.yaml

Configure atmos.yaml:

stacks:
  included_paths: ["orgs/**/*"]
  excluded_paths: ["**/_defaults.yaml"]
  name_template: "{{.vars.tenant}}-{{.vars.stage}}"

Layered Stack Configuration

Groups components by infrastructure function (network, data, compute). Each layer imports its relevant catalog defaults. Different teams can own different layers. Environments import only the layers they need.

# stacks/layers/network.yaml
import:
  - catalog/vpc/defaults
# stacks/layers/data.yaml
import:
  - catalog/rds/defaults

# stacks/deploy/prod.yaml
import:
  - layers/network
  - layers/data
  - layers/compute
vars:
  stage: prod

The _defaults.yaml Convention

A naming convention (not an Atmos feature) for organizing hierarchical defaults:

• Underscore prefix ensures files sort to top of directory listings
• Excluded from stack discovery via excluded_paths: ["**/_defaults.yaml"]
• Must be explicitly imported -- Atmos does NOT auto-import them
• Creates clear inheritance chains when each level imports its parent

Best practices: keep to 3-4 levels maximum, document import chains, use base-relative paths (resolved from stacks.base_path).

Configuration Catalog Patterns

Basic Catalog

Mirror your component directory in stacks/catalog/. Each component gets a defaults.yaml with shared configuration.

stacks/catalog/
  vpc/
    defaults.yaml          # Base defaults for all VPCs
    dev.yaml               # Dev-specific overrides
    prod.yaml              # Prod-specific overrides
    ue2.yaml               # Region-specific (imports defaults)
  s3-bucket/
    defaults.yaml
    public.yaml            # Archetype: public website bucket
    logging.yaml           # Archetype: log storage bucket
    artifacts.yaml         # Archetype: CI/CD artifacts

Mixins

Reusable configuration fragments that encapsulate settings applied consistently across stacks. Two scopes:

Global mixins (stacks/mixins/) -- region defaults, stage defaults, tenant defaults:

# stacks/mixins/region/us-east-2.yaml
vars:
  region: us-east-2
  environment: ue2
components:
  terraform:
    vpc:
      vars:
        availability_zones: [us-east-2a, us-east-2b, us-east-2c]

Catalog mixins (stacks/catalog/<component>/mixins/) -- feature flags, versions:

# stacks/catalog/eks/mixins/1.28.yaml
components:
  terraform:
    eks/cluster:
      vars:
        cluster_kubernetes_version: "1.28"
        addons:
          vpc-cni:
            addon_version: "v1.14.1-eksbuild.1"

Import order matters -- later imports override earlier ones. Order from general to specific:

import:
  - catalog/vpc/defaults         # 1. Component defaults
  - catalog/vpc/mixins/multi-az  # 2. Feature flags
  - mixins/region/us-east-2      # 3. Region settings
  - mixins/stage/prod            # 4. Stage settings (most specific)

Component Archetypes

Pre-configured variants for specific use cases. Define abstract base components with metadata.type: abstract, then create archetypes that inherit from the base with use-case-specific settings.

Catalog Templates

Use Go templates in imports to dynamically generate component instances. Import the same template multiple times with different context values:

import:
  - path: catalog/eks/iam-role/defaults.tmpl
    context:
      app_name: "auth"
      service_account_name: "auth"
      service_account_namespace: "auth"

Use sparingly -- the templating engine is powerful but can reduce maintainability.

Inheritance Patterns

Component Inheritance

A component inherits configuration from a base using metadata.inherits:

components:
  terraform:
    vpc:
      metadata:
        component: vpc
        inherits:
          - vpc/defaults    # Inherit all vars, then override
      vars:
        max_subnet_count: 2

Inheritance order: base component -> inherited components (in order) -> inline vars.

Abstract Components

Mark components as non-deployable blueprints with metadata.type: abstract. Prevents accidental atmos terraform apply on base configurations. Components inheriting from abstract bases are deployable by default.

# In catalog
vpc/defaults:
  metadata:
    type: abstract
  vars:
    enabled: true
    nat_gateway_enabled: true

Multiple Component Instances

Deploy multiple instances of the same Terraform component in one environment by defining multiple Atmos components pointing to the same metadata.component:

components:
  terraform:
    vpc/1:
      metadata:
        component: vpc
        inherits: [vpc/defaults]
      vars:
        name: vpc-1
        ipv4_primary_cidr_block: 10.9.0.0/18
    vpc/2:
      metadata:
        component: vpc
        inherits: [vpc/defaults]
      vars:
        name: vpc-2
        ipv4_primary_cidr_block: 10.10.0.0/18

Multiple Inheritance

Inherit from multiple abstract bases to compose configuration from independent concerns:

rds:
  metadata:
    component: rds
    inherits:
      - base/defaults     # Applied first
      - base/logging      # Applied second
      - base/production   # Applied last, highest precedence

Merge behavior: scalars -- later wins; maps -- deep merged; lists -- later replaces entirely.

Configuration Composition

Inline Configuration

Define components directly in stack manifests. Use for prototyping, single-environment deployments, or components unique to one stack.

Partial Component Configuration

Split a component's configuration across multiple files imported into the same stack. Useful for independently managing parts of complex configurations (e.g., EKS cluster defaults + Kubernetes version mixin).

Component Overrides

Apply configuration to a subset of components without affecting others using the overrides section. Overrides are file-scoped and do not get inherited.

# stacks/teams/platform.yaml
import:
  - catalog/vpc/defaults
  - catalog/eks/defaults
terraform:
  overrides:
    vars:
      tags:
        Team: Platform     # Only applies to vpc and eks, not other teams' components

DRY Configuration with Locals

File-scoped variables that reduce repetition within a single stack file:

locals:
  prefix: "{{ .locals.namespace }}-{{ .locals.environment }}"
components:
  terraform:
    vpc:
      vars:
        name: "{{ .locals.prefix }}-vpc"

Locals are not inherited across imports. Use vars or settings for cross-file values.

Version Management Patterns

Continuous Version Deployment (Recommended)

Trunk-based strategy where all environments reference the same component path and converge through progressive automated rollout. Simplest approach with strongest feedback loops.

Folder-Based Versioning

Components organized in explicit folders (vpc/v1/, vpc/v2/). Environments reference specific version folders. Use metadata.name for stable workspace keys across version upgrades.

vpc:
  metadata:
    name: vpc            # Stable identity (workspace key stays same)
    component: vpc/v2    # Version can change freely

Release Tracks/Channels

Named channels (alpha/vpc, beta/vpc, prod/vpc) that environments subscribe to. Promote tracks instead of individual environment pins. Use label-based versioning schemes (maturity levels, environment names).

Strict Version Pinning

Explicit SemVer versions (vpc/1.2.3). Works with vendoring from external sources. Use number-based versioning schemes. Higher operational overhead but strongest audit trail.

Source-Based Version Pinning

Per-environment version control using the source field in stack configuration. Just-in-time vendoring without managing separate vendor manifests.

vpc:
  source:
    uri: github.com/org/components//modules/vpc
    version: 1.450.0

Vendoring Component Versions

Automate copying components from external sources with vendor.yaml and atmos vendor pull. Provides local control, audit trail, and searchable codebase. Complements any deployment strategy.

Git Flow: Branches as Channels

Branch-based alternative where long-lived branches map to release channels. Promotions happen via merges. Best for teams already practicing Git Flow.

When to Use Which Pattern

Anti-Patterns to Avoid

• Vendoring multiple versions to the same path -- last one overwrites all previous
• Including version in workspace_key_prefix -- breaks state continuity during upgrades
• Mixing trunk-based and Git Flow -- creates team confusion about promotion paths
• Over-pinning environments -- creates high operational overhead and weak feedback loops
• Inconsistent path conventions -- pick {track}/{component} or {component}/{track} and stick with it
• Assuming _defaults.yaml auto-imports -- they must always be explicitly imported
• Too many inheritance levels -- keep to 3-4 levels maximum for maintainability

References

For detailed examples and directory layouts, see:

• references/stack-organization.md -- Stack organization patterns with directory layouts
• references/version-management.md -- Versioning strategies with implementation details