backend-design（后端设计）

<domain_overview>

Backend Design System

Philosophy: The Backend is the Fortress. Logic is Law. Latency is the Enemy. Core Principle: ISOLATE features. TRUST no one. SCALE linearly.

ANTI-HAPPY PATH MANDATE (CRITICAL): Never assume the ideal scenario. AI-generated code often fails by ignoring edge cases and failure modes. For every business logic slice, you MUST document and test at least three failure scenarios: Race Conditions, Data Integrity violations (e.g., unique constraint overlaps), and Boundary failures. Reject any implementation that only covers the 'Happy Path'. Engineering is the art of handling what shouldn't happen. </domain_overview>

<architectural_protocols>

🚀 ELITE TIER KNOWLEDGE (ARCHITECTURAL PROTOCOLS)

0. The "Vertical Slice" Law (The Anti-Layer Mandate)

CRITICAL: You are FORBIDDEN from creating "Horizontal Layers" (Controllers, Services, Repositories) as primary folders.

The "Feature-First" Protocol: Code must be organized by BUSINESS CAPABILITY, not technical role.

1. The Slice: A single directory (e.g., features/create-order/) contains EVERYTHING needed for that feature:

* handler.ts (Controller) * logic.ts (Domain/Service) * schema.ts (DTO/Validation) * db.ts (Data Access)

1. The Benefit: Changing a feature requires touching only ONE folder. No "Shotgun Surgery" across 5 layers.
2. Shared Kernel: Only truly generic code (Logging, Auth Middleware, Database Connection) goes into shared/.

1. The "Modular Monolith" Mandate

• Microservices Ban: Do NOT start with microservices. Start with a Modular Monolith.
• Modulith Rules:

* Modules must be isolated (like internal microservices). * Modules communicate via Events (Sub-Process or Message Bus), NEVER by importing another module's code directly. * The Outbox Pattern (Guaranteed Delivery): * *Problem:* If DB commit succeeds but Event Bus fails, the system is inconsistent. * *Mandate:* Write events to an outbox table in the SAME transaction as the data change. * *Relay:* A background worker pushes outbox entries to the Message Bus (RabbitMQ/Kafka). * Data Sovereignty: Module A cannot query Module B's tables. It must ask Module B via API/Event.

2. The "Zero Trust" Security Protocol

Detailed protocols: See security-protocols.md

Quick Rules:

1. Strict Serialization: NEVER return raw DB entities → Use ResponseDTO
2. Validation at Gate: Schema validation (Zod/Pydantic) BEFORE logic
3. Token Sovereignty: PASETO v4 > JWT (Ed25519 if JWT forced)

</architectural_protocols>

<reliability_contracts>

🏗️ Reliability & Performance Contracts

3. The "Sub-100ms" Performance Mandate

• The Latency Budget: P50 < 100ms. P99 < 500ms.
• UUIDv7 (The Time-Lord Rule):

* *Ban:* Never use UUIDv4 (Random) for Primary Keys. It fragments B-Tree indexes. * *Mandate:* Use UUIDv7 (Time-ordered). It enables clustered index locality (fast inserts) like integers, with the uniqueness of UUIDs.

• N+1 Assassin:

* *Check:* Always inspect ORM queries. Loops triggering DB calls are a "Level 0" error. * *Fix:* Use DataLoader pattern or explicit JOIN loading.

4. API Reliability Contracts

• RFC 7807 (Problem Details):

* *Ban:* returning { "error": "Something went wrong" }. * *Mandate:* Return standard Problem JSON:

        {
          "type": "https://api.myapp.com/errors/insufficient-funds",
          "title": "Insufficient Funds",
          "status": 403,
          "detail": "Current balance is 10.00, required is 15.00",
          "instance": "/transactions/12345"
        }

• Idempotency Keys:

* *Rule:* All critical POST/PATCH (Money, State Change) must accept an Idempotency-Key header. * *Logic:* If key exists in Cache (24h TTL), return stored response without re-executing logic. </reliability_contracts>

<database_integrity>

🗄️ Database Integrity & Design

5. Database Integrity & Design

• Hard Constraints: Application-level checks are "Suggestions". Database Constraints (Foreign Keys, Unique Indexes, Check Constraints) are "Laws".
• Cursor Pagination:

* *Ban:* OFFSET / LIMIT on large tables (O(N) performance degradation). * *Mandate:* Cursor-based pagination (WHERE created_at < cursor LIMIT 20).

• Migration Discipline:

* Never alter a column in a way that locks the table for >1s. * Use "Expand and Contract" pattern for breaking changes.

• Concurrency Control:

* *Problem:* Two users update the same record. The last one wipes the first. * *Mandate:* Use Optimistic Locking. Add a version (int) column. * *Logic:* Update WHERE id = X AND version = Y. If 0 rows affected, throw StaleObjectException.

6. AI & Vector Readiness

• Semantic Storage: Backend must be ready to store embeddings (Vector Types).
• Guardrails: Output from LLMs must be sanitized and structure-checked on the server side before returning to frontend.

</database_integrity>

<observability>

👁️ Observability & Monitoring (The "Glass Box" Protocol)

7. Structured Logging Only

• Ban: console.log("User updated"). String logs are useless for machines.
• **Mandate:* JSON Logs with correlation IDs. { "level": "info", "event": "user_updated", "user_id": "u7-...", "trace_id": "..." }.

8. Distributed Tracing (OpenTelemetry)

• Every request MUST carry a traceparent header.
• Spans must cover: DB Queries, External API Calls, and Redis operations.

9. Health Checks

• Liveness (/health/live): "Am I running?" (Instant, no checks).
• Readiness (/health/ready): "Can I take traffic?" (Check DB/Redis connection).

</observability>

<resilience>

🛡️ Resilience Patterns (The "Anti-Fragile" Mandate)

10. Circuit Breakers

• Wrap ALL external calls (Payment Gateways, 3rd Party APIs) in a Circuit Breaker.
• *Logic:* After 5 failures, fail fast for 30s. Don't drown the downstream service.

11. Rate Limiting

• Protect *every* public endpoint with a Token Bucket rate limiter (Redis-backed).
• Differentiate limits by User Role (Anon: 60/min, Pro: 1000/min).

</resilience>

<workflow_rules>

🔧 Workflow Rules

1. The Pre-Flight Checklist

1. Environment Hardening:

* Verify all process.env variables at startup using a schema (e.g., t3-env or envalid). If a key is missing, crash immediately. Do not start the server in an undefined state. Before writing a single handler:

1. Define the DTOs: Request Schema (Zod) and Response Schema.
2. Define the Error States: What can go wrong? (404, 409, 429).
3. Define the Data Access: What is the most efficient SQL query?

2. The "No Magic" Rule

• Avoid "Magical" ORM features (Lazy Loading, Auto-Saving context).
• Prefer Explicit over Implicit. "Write the SQL (or Query Builder) if the ORM hides expensive logic."

3. Testing Pyramid

1. Unit: Test Domain Logic in isolation (mock DB).
2. Integration: Test Feature Slice with a REAL containerized DB (Testcontainers).
3. E2E: Test critical flows from the "Outside".

</workflow_rules>

<audit_and_reference>

📂 Cognitive Audit Cycle

Before committing code:

1. Is the endpoint under a feature slice? (Not in a generic controller folder).
2. Is Input Validated with a Schema? (Zero Trust).
3. Are DB Indexes used? (Run EXPLAIN ANALYZE).
4. Is the Primary Key UUIDv7? (Index Perf).
5. Are secrets managed properly? (No hardcoded strings).

🔗 CROSS-SKILL INTEGRATION

Command: Use these skills to architect "Fortress-Level" backend systems.

</audit_and_reference>