system-architecture（系统架构）

System Architecture

When to Use

Activate this skill when:

• Designing a new module, service, or major feature that requires structural decisions
• Choosing between architectural approaches (e.g., where to place logic, how to structure data flow)
• Planning database schema changes or refactoring existing schema
• Making frontend state management decisions (server state vs client state, context vs store)
• Evaluating technology trade-offs for a new capability
• Creating or reviewing Architecture Decision Records (ADRs)
• Setting up a new project or major subsystem from scratch

Input: If plan.md exists (from project-planner), read it for context about the feature scope and affected modules. Otherwise, work from the user's request directly.

Output: Write architecture decisions to architecture.md and create ADRs in docs/adr/ADR-NNN-<title>.md. Tell the user: "Architecture written to architecture.md. Run /api-design-patterns for API contracts or /task-decomposition for implementation tasks."

Do NOT use this skill for:

• Writing implementation code (use python-backend-expert or react-frontend-expert)
• API contract design or endpoint specifications (use api-design-patterns)
• Testing patterns or strategies (use pytest-patterns or react-testing-patterns)
• Deployment or infrastructure decisions (use docker-best-practices or deployment-pipeline)

Instructions

Project Layer Architecture

The standard Python/React full-stack architecture follows a layered pattern with strict dependency direction.

Backend Layers (FastAPI)

HTTP Request
    ↓
┌─────────────────────┐
│   Routers (routes/)  │  ← HTTP concerns: request parsing, response formatting, status codes
│                      │     Uses: Depends() for injection, Pydantic schemas for validation
├─────────────────────┤
│   Services           │  ← Business logic: orchestration, validation rules, domain operations
│   (services/)        │     No HTTP awareness. Raises domain exceptions, not HTTPException.
├─────────────────────┤
│   Repositories       │  ← Data access: queries, CRUD operations, database interactions
│   (repositories/)    │     No business logic. Returns model instances or None.
├─────────────────────┤
│   Models (models/)   │  ← SQLAlchemy ORM models: table definitions, relationships, indexes
│   Schemas (schemas/) │  ← Pydantic v2 models: request/response contracts, validation
└─────────────────────┘
    ↓
Database

Dependency direction rules:

• Routers depend on Services (never on Repositories directly)
• Services depend on Repositories (never on Routers)
• Repositories depend on Models (never on Services)
• Schemas are shared across layers but define no dependencies themselves
• Never skip layers: no direct database access from routes

Dependency injection pattern:

# Router depends on Service via Depends()
@router.post("/users", response_model=UserResponse)
async def create_user(
    data: UserCreate,
    service: UserService = Depends(get_user_service),
) -> UserResponse:
    return await service.create_user(data)

# Service depends on Repository via constructor injection
class UserService:
    def __init__(self, repo: UserRepository) -> None:
        self.repo = repo

# Repository depends on AsyncSession via Depends()
class UserRepository:
    def __init__(self, session: AsyncSession) -> None:
        self.session = session

Frontend Layers (React/TypeScript)

┌─────────────────────┐
│   Pages (pages/)     │  ← Route-level components: data fetching, layout composition
├─────────────────────┤
│   Layouts            │  ← Page structure: navigation, sidebars, content areas
│   (layouts/)         │
├─────────────────────┤
│   Features           │  ← Domain-specific: UserProfile, OrderList, ChatPanel
│   (features/)        │     Composed from shared components + hooks
├─────────────────────┤
│   Shared Components  │  ← Reusable UI: Button, Modal, Table, Form, Input
│   (components/)      │     No business logic. Configurable via props.
├─────────────────────┤
│   Hooks (hooks/)     │  ← Custom hooks: useAuth, usePagination, useDebounce
│   API (api/)         │  ← API client functions, TanStack Query configurations
├─────────────────────┤
│   Types (types/)     │  ← Shared TypeScript interfaces and type definitions
└─────────────────────┘

Component dependency direction:

• Pages import Features and Layouts
• Features import Shared Components and Hooks
• Shared Components import only other Shared Components and Types
• Hooks import API functions and Types
• API functions import Types only

Decision Framework

When facing architectural decisions, follow this structured process:

Step 1: Define the Problem

• What capability is needed?
• What are the non-functional requirements? (performance, scalability, maintainability)
• What constraints exist? (team size, timeline, existing infrastructure)

Step 2: Identify Options

• List 2-3 viable architectural approaches
• For each option, document:

- How it works (brief technical description) - Advantages - Disadvantages - Risks

Step 3: Evaluate Against Criteria

Step 4: Decide and Document

• Choose the option that best satisfies the weighted criteria
• Document the decision in an ADR (see references/architecture-decision-record-template.md)
• Record what was NOT chosen and why — this context is valuable for future decisions

Step 5: Communicate

• Share the ADR with the team
• Identify any migration or rollout steps needed
• Flag reversibility: is this a one-way door or a two-way door?

Database Schema Design

Design Principles

1. Start normalized (3NF) — Denormalize only for proven performance bottlenecks, not speculation
2. One migration per logical change — Each Alembic migration should represent a single, coherent schema modification
3. Always include downgrade — Every migration must have a working downgrade() function
4. Index strategically:

- Primary keys (automatic) - Foreign keys (always) - Columns in WHERE clauses of frequent queries - Composite indexes for multi-column lookups - Partial indexes for filtered queries (e.g., WHERE is_active = true)

SQLAlchemy 2.0 Async Patterns

# Model definition with Mapped types (SQLAlchemy 2.0 style)
class User(Base):
    __tablename__ = "users"

    id: Mapped[int] = mapped_column(primary_key=True)
    email: Mapped[str] = mapped_column(String(255), unique=True, index=True)
    is_active: Mapped[bool] = mapped_column(default=True)
    created_at: Mapped[datetime] = mapped_column(server_default=func.now())

    # Relationships: ALWAYS use eager loading with async
    posts: Mapped[list["Post"]] = relationship(
        back_populates="author",
        lazy="selectin",  # or "joined" — NEVER "lazy" with async
    )

Async session rules:

• One AsyncSession per request — never share across concurrent tasks
• Use async with context manager for automatic cleanup
• Map session boundaries to transaction boundaries
• Use selectin or joined loading — lazy loading is incompatible with asyncio
• Use run_sync() only as a last resort for legacy code

Migration Planning

1. Schema change → Generate migration: alembic revision --autogenerate -m "description"
2. Review generated migration — verify column types, indexes, constraints
3. Test upgrade: alembic upgrade head
4. Test downgrade: alembic downgrade -1
5. Test data preservation: ensure existing data survives the round-trip

Frontend Architecture

State Management Decision Tree

Is the data from the server?
├── YES → Use TanStack Query (useQuery, useMutation)
│         Configure staleTime, gcTime, query keys
│
└── NO → Is it needed across multiple components?
         ├── YES → Is it complex with actions/reducers?
         │         ├── YES → Use Zustand store
         │         └── NO  → Use React Context
         │
         └── NO → Use useState / useReducer locally

TanStack Query conventions:

• Query keys: [resource,...identifiers] (e.g., ["users", userId], ["posts", {page, limit}])
• Use queryOptions() factory to centralize key + fn definitions — prevents copy-paste key errors
• Set staleTime based on data freshness needs (default 0 is too aggressive for most cases)
• Invalidate with invalidateQueries() after mutations — never manual refetch()
• Handle all states: isPending, isError, data

Component design rules:

• Props for configuration, hooks for data
• Lift state only as high as needed — no premature context creation
• Keep components under 200 lines — extract sub-components or custom hooks when larger
• Use children and composition over deep prop drilling

Routing Structure

Organize routes to mirror the URL structure:

src/
├── pages/
│   ├── HomePage.tsx           → /
│   ├── LoginPage.tsx          → /login
│   ├── users/
│   │   ├── UserListPage.tsx   → /users
│   │   └── UserDetailPage.tsx → /users/:id
│   └── settings/
│       └── SettingsPage.tsx   → /settings

Cross-Cutting Concerns

Authentication Flow

Login Request
    ↓
Backend: Validate credentials → Generate JWT (access + refresh tokens)
    ↓
Frontend: Store access token in memory, refresh token in httpOnly cookie
    ↓
API Calls: Attach access token via Authorization header
    ↓
Token Expired: Use refresh token to obtain new access token
    ↓
Refresh Failed: Redirect to login

Architecture decisions for auth:

• Access tokens: short-lived (15-30 min), stored in memory (not localStorage)
• Refresh tokens: longer-lived (7-30 days), stored in httpOnly cookie
• Backend: FastAPI Depends() chain for token validation → user extraction → permission check
• Frontend: Auth context providing user, login(), logout(), isAuthenticated

Error Handling Strategy

Errors should be handled at the appropriate layer:

Backend exception hierarchy:

class AppError(Exception):
    """Base application error."""

class NotFoundError(AppError):
    """Resource not found."""

class ConflictError(AppError):
    """Resource conflict (duplicate, version mismatch)."""

class ValidationError(AppError):
    """Business rule violation."""

Router-level exception handler maps domain exceptions to HTTP responses:

@app.exception_handler(NotFoundError)
async def not_found_handler(request: Request, exc: NotFoundError):
    return JSONResponse(status_code=404, content={"detail": str(exc)})

Logging Architecture

Backend (structlog):

• Structured JSON logs in production
• Human-readable console in development
• Bind request context (request_id, user_id) at middleware level
• Log at service layer (business events), not repository layer (too noisy)
• Use log levels: DEBUG (development only), INFO (business events), WARNING (recoverable issues), ERROR (failures requiring attention)

Frontend:

• console.* in development
• Structured error reporting to backend or Sentry in production
• Log user actions for debugging, not for analytics

Configuration Management

Backend (pydantic-settings):

class Settings(BaseSettings):
    model_config = SettingsConfigDict(env_file=".env")

    database_url: str
    redis_url: str = "redis://localhost:6379"
    jwt_secret: str
    debug: bool = False

Frontend (environment variables):

• VITE_API_URL for API base URL
• Build-time injection via Vite's import.meta.env
• No secrets in frontend environment variables

Output Files

architecture.md

Write the architecture document to architecture.md at the project root:

# Architecture: [Feature/System Name]

## Overview
[1-2 sentence summary of the architectural approach]

## Layer Structure
[Backend and frontend layer descriptions from this skill's patterns]

## Key Decisions
[Summary of decisions made, with links to ADRs]

## Database Schema
[Entity descriptions, relationships, key indexes]

## Cross-Cutting Concerns
[Auth, error handling, logging approach]

## Next Steps
- Run `/api-design-patterns` to define API contracts
- Run `/task-decomposition` to create implementation tasks

ADRs

For each significant decision, create an ADR in docs/adr/:

# ADR-NNN: [Decision Title]

## Status
Accepted | Proposed | Superseded

## Context
[Why this decision is needed]

## Decision
[What we decided]

## Consequences
[Positive and negative outcomes]

Number ADRs sequentially (ADR-001, ADR-002, etc.).

Examples

Architecture Decision: Real-Time Notifications

Problem: The application needs real-time notifications for users (new messages, status updates).

Options evaluated:

Decision: WebSocket for this use case.

Rationale: Notifications require low latency and the system will eventually need bidirectional communication (typing indicators, presence). SSE would work for notifications alone but would require a separate solution for future bidirectional needs. Polling introduces unacceptable latency for real-time UX.

Architecture:

• Backend: FastAPI WebSocket endpoint with ConnectionManager class
• Frontend: Custom useWebSocket hook with automatic reconnection
• Scaling: Redis pub/sub for multi-instance message distribution
• Persistence: Store notifications in database for offline users
• Fallback: REST endpoint for notification history and initial load

See references/architecture-decision-record-template.md for the full ADR format.

Edge Cases

Monolith vs Microservices

Default to modular monolith for teams smaller than 10 developers. A modular monolith provides:

• Clear module boundaries without network overhead
• Shared database with module-specific schemas
• Easy refactoring and code navigation
• Simple deployment and debugging

Consider microservices only when:

• Independent scaling is required for specific components
• Different modules need different technology stacks
• Team size exceeds 10 and ownership boundaries are clear
• Deployment independence is a business requirement

Migration path: Design module boundaries in the monolith as if they were services (no direct cross-module database access, communicate via service interfaces). This makes extraction to microservices straightforward when needed.

When to Break the Layer Pattern

The strict Router → Service → Repository pattern should be followed for standard CRUD operations. Acceptable exceptions:

• Background tasks: May call services directly without going through a router
• Event handlers: Domain event listeners may call services from any context
• CLI commands: Management scripts may access services or repositories directly
• Migrations: Data migrations may access models directly (no service/repo layer needed)
• Health checks: May access the database directly for simple connectivity verification

In all cases, business logic should still live in the service layer — these exceptions are about the entry point, not about bypassing business rules.

Evolving Architecture

When the architecture needs to change:

1. Write an ADR documenting the motivation and the proposed change
2. Identify all affected modules and their dependencies
3. Plan an incremental migration — never big-bang rewrites
4. Maintain backward compatibility during transition (strangler fig pattern)
5. Set a deadline for completing the migration and removing legacy code