idempotency-patterns（幂等性模式）

Idempotency Patterns

Patterns for designing APIs and systems that handle retries safely without duplicate side effects.

When to Use This Skill

• Designing APIs that handle retries safely
• Implementing idempotency keys
• Preventing duplicate operations
• Building reliable payment/order systems
• Handling network failures gracefully

What is Idempotency?

Idempotent operation: Same result regardless of how many times executed

f(x) = f(f(x)) = f(f(f(x))) = ...

Examples:
- GET /user/123      → Always returns same user (idempotent)
- DELETE /user/123   → User deleted once, subsequent calls no-op (idempotent)
- POST /orders       → Creates new order each time (NOT idempotent)

Why Idempotency Matters

Network reality:
Client ──request──> Server
       <──response── (lost!)

Client doesn't know if request succeeded.
Should it retry?

Without idempotency:
- Retry creates duplicate order
- Customer charged twice
- Inventory decremented twice

With idempotency:
- Retry returns same result
- No duplicate side effects
- Safe to retry

HTTP Method Idempotency

Idempotency Key Pattern

Concept

Client generates unique key, server tracks processed keys

Request 1:
POST /payments
Idempotency-Key: abc-123
{amount: 100}
→ Process payment, store result with key abc-123

Request 2 (retry):
POST /payments
Idempotency-Key: abc-123
{amount: 100}
→ Find stored result for abc-123, return same response
→ No duplicate payment

Implementation

Idempotency store schema:
┌──────────────────────────────────────────────────┐
│ idempotency_key │ request_hash │ response │ ttl │
├──────────────────────────────────────────────────┤
│ abc-123         │ sha256(...)  │ {...}    │ 24h │
└──────────────────────────────────────────────────┘

Flow:
1. Receive request with idempotency key
2. Check if key exists in store
3. If exists:
   a. Verify request_hash matches (same request)
   b. Return stored response
4. If not exists:
   a. Process request
   b. Store response with key
   c. Return response

Key Generation

Client-generated keys (recommended):
- UUID v4: 550e8400-e29b-41d4-a716-446655440000
- ULID: 01ARZ3NDEKTSV4RRFFQ69G5FAV
- Custom: {client_id}-{timestamp}-{random}

Requirements:
- Globally unique
- Unpredictable (prevent guessing)
- Client controls key

Request Fingerprinting

Verify retry is same request (not just same key):

request_hash = hash(
  method,
  path,
  body,
  relevant_headers
)

If idempotency_key exists but request_hash differs:
→ Return 422: "Idempotency key reused with different request"

At-Most-Once vs At-Least-Once

At-Most-Once

Operation executes 0 or 1 time, never more.

Use when: Duplicate is worse than missing
- Payment processing
- Order creation
- Resource provisioning

Implementation: Idempotency keys with deduplication

At-Least-Once

Operation executes 1 or more times.

Use when: Missing is worse than duplicate
- Event notifications
- Log ingestion
- Analytics events

Implementation: Retry until acknowledged, handle duplicates downstream

Exactly-Once (Hard)

Operation executes exactly 1 time.

Extremely difficult in distributed systems.
Usually achieved through:
- At-least-once delivery + idempotent processing
- Distributed transactions (2PC)
- Saga pattern with compensation

Duplicate Detection Strategies

Strategy 1: Idempotency Key Store

Store: Redis or database

Key: idempotency_key
Value: {
  status: "processing" | "completed" | "failed",
  response: {...},
  created_at: timestamp,
  expires_at: timestamp
}

TTL: 24-72 hours typically

Strategy 2: Natural Key Deduplication

Use business identifiers:
- Order: {customer_id}-{cart_id}-{timestamp}
- Payment: {order_id}-{amount}-{currency}
- Transfer: {sender}-{receiver}-{reference}

Check if natural key exists before processing.

Strategy 3: Database Constraints

CREATE TABLE orders (
  id UUID PRIMARY KEY,
  idempotency_key VARCHAR(255) UNIQUE,
  ...
);

INSERT fails if idempotency_key already exists.

Strategy 4: Optimistic Locking

UPDATE accounts
SET balance = balance - 100, version = version + 1
WHERE id = 123 AND version = 5;

If version changed, retry with new version.
Prevents concurrent duplicate updates.

Handling In-Flight Requests

Problem: Request A starts, Request B (retry) arrives before A completes

Solution 1: Lock on idempotency key
- First request acquires lock
- Retry waits or returns "processing"

Solution 2: Status tracking
- Store "processing" status immediately
- Retry sees "processing", waits or returns 409

Response for in-flight:
HTTP 409 Conflict
{
  "error": "Request with this idempotency key is still processing",
  "retry_after": 5
}

Idempotency in Different Contexts

Payment APIs

POST /charges
Idempotency-Key: {uuid}
{
  "amount": 1000,
  "currency": "usd",
  "source": "tok_visa"
}

Critical: Never charge twice
Store: idempotency_key → charge_id, status, response
TTL: 24-48 hours

Message Queues

Producer:
- Include message_id in payload
- Retry with same message_id

Consumer:
- Track processed message_ids
- Skip if already processed

Deduplication window: Based on expected retry window

Database Operations

Insert with idempotency:
INSERT INTO orders (id, idempotency_key, ...)
VALUES (gen_id(), 'abc-123', ...)
ON CONFLICT (idempotency_key) DO NOTHING
RETURNING *;

If conflict, fetch existing record.

Event Sourcing

Events naturally idempotent by sequence:
- Event ID: {aggregate_id}-{sequence_number}
- Reject if sequence already exists
- Replay is safe (events are immutable)

Best Practices

Key Storage

- Use fast store (Redis) for hot path
- Persist to database for durability
- Set appropriate TTL (24-72 hours typical)
- Clean up expired keys

Error Handling

If processing fails:
1. Store failure response with key
2. Client retries get same error
3. Client must use NEW key to try again

This prevents infinite retry loops on bad requests.

Documentation

Document clearly:
- Which endpoints require idempotency keys
- Key format requirements
- TTL for stored results
- Error responses for duplicates

Client Implementation

1. Generate idempotency key before first attempt
2. Store key locally until confirmed success
3. Retry with SAME key on network failure
4. Generate NEW key for genuinely new requests
5. Don't reuse keys across different operations

Common Pitfalls

1. Storing only success responses
   → Store failures too, otherwise retry creates duplicate

2. Short TTL
   → Client might retry after TTL expires, causing duplicate

3. Not hashing request body
   → Different requests with same key processed differently

4. Race conditions on concurrent retries
   → Use locks or atomic operations

5. Not handling partial failures
   → Use sagas or compensation for multi-step operations

Related Skills

• api-design-fundamentals - API design patterns
• rate-limiting-patterns - Handling retries
• distributed-transactions - Multi-step operations