Estrategia de Arquitectura de Caché

# Caching Architecture Plan ## Context - **Performance Requirements**: The application requires sub-10ms read latency for hot data paths and p99 latency under 50ms. Current database latency averages 25ms (p50) and 80ms (p99), failing to meet SLAs. - **Bottlenecks**: Database CPU saturation at 70% during peak loads; repetitive queries for identical data; slow response times for frequently accessed user profiles and product catalog items. - **Access Patterns**: Read-heavy workload (95% reads / 5% writes) with high temporal locality on user sessions and product recommendations. Hot keys represent ~20% of data but receive ~80% of traffic. - **Consistency Requirements**: User profile data requires strong consistency (max staleness: 0s). Product inventory requires eventual consistency (max staleness: 60s). Session data requires eventual consistency (max staleness: 5 minutes). - **Infrastructure Constraints**: Existing AWS deployment with 3 application instances. Current caching: none in production. Database is PostgreSQL 14 (primary + 2 read replicas). Network latency between services: <1ms. ## Caching Architecture Plan - [ ] **CACHE-PLAN-1.1 [CDN Layer Design]**: - **Layer**: CDN - **Technology**: Amazon CloudFront - **Scope**: Static assets (images, CSS, JS), public product catalog pages (read-only) - **Configuration**: TTL 24h for static assets, TTL 5m for catalog pages; use `Cache-Control: max-age=86400, stale-while-revalidate=86400` headers; enable compression (Brotli) and HTTP/2 - [ ] **CACHE-PLAN-1.2 [Application Cache Layer Design]**: - **Layer**: Application (in-process) - **Technology**: Caffeine (Java) - version 3.1.8 - **Scope**: Highly accessed user session data (hot keys), frequently used configuration data - **Configuration**: Maximum size 10,000 entries per instance; TTL 5 minutes; eviction policy: LFU with maximum size; local caching only (no replication) - [ ] **CACHE-PLAN-1.3 [Distributed Cache Layer Design]**: - **Layer**: Distributed - **Technology**: Redis 7.0 (ElastiCache, cluster mode enabled) - **Scope**: User profile data, product catalog data, inventory counts, API response cache for aggregated endpoints - **Configuration**: 3 shards, each with 1 replica; `maxmemory-policy: allkeys-lfu`; TTL per key type (see CACHE-ITEM sections); connection pooling with 50 connections per instance; TLS enabled; enable `notify-keyspace-events Ex` for expiration events - [ ] **CACHE-PLAN-1.4 [Database Cache Layer Design]**: - **Layer**: Database - **Technology**: PostgreSQL 14 (read replicas) with statement-level caching - **Scope**: Frequently executed join queries (product with category), aggregate queries (inventory levels) - **Configuration**: 2 read replicas; circuit breaker for replica routing (fallback to primary); no TTL; invalidation via materialized view refresh every 30 seconds ## Caching Items - [ ] **CACHE-ITEM-1.1 [User Profile Cache Implementation]**: - **Description**: Cache user profile data in Redis to reduce database load. Key: `user:profile:{userId}`. Value: JSON-serialized profile (Protocol Buffers format). - **Invalidation Strategy**: Write-through (update cache synchronously on profile change) - **TTL and Eviction**: TTL 24h; eviction policy `allkeys-lfu`; jittered TTL (add ±30s random variation) - **Validation**: Compare database and cache responses for 100 random users; assert equality and check p50/p99 latency against SLAs - [ ] **CACHE-ITEM-1.2 [Product Catalog Cache Implementation]**: - **Description**: Cache product catalog data (name, description, image URL) in Redis with CDN edge caching. Key: `product:catalog:{productId}`. Value: Protocol Buffer binary. - **Invalidation Strategy**: Event-driven (publish invalidation event to Redis Stream on product update) - **TTL and Eviction**: TTL 1h for Redis; TTL 24h for CDN; eviction policy `allkeys-lfu`; jittered TTL (add ±1 minute) - **Validation**: Trigger product update API, verify CDN cache is purged within 30 seconds and Redis key is invalidated within 1 second - [ ] **CACHE-ITEM-1.3 [Session Data Cache Implementation]**: - **Description**: Cache user session data in Caffeine (local cache) with Redis as fallback. Key: `session:{sessionToken}`. Value: Serialized session object (MessagePack). - **Invalidation Strategy**: Cache-aside (load from cache, fallback to database on miss); TTL-based expiration - **TTL and Eviction**: TTL 5 minutes for local cache; TTL 30 minutes for Redis; LFU eviction for local; `volatile-ttl` for Redis expired keys - **Validation**: Simulate session expiry and verify user is redirected to login; test cross-instance session persistence via Redis fallback - [ ] **CACHE-ITEM-1.4 [Inventory Count Cache Implementation]**: - **Description**: Cache inventory counts for product detail pages. Key: `inventory:count:{productId}`. - **Invalidation Strategy**: Write-behind (asynchronously update cache after database write; publish event on change) - **TTL and Eviction**: TTL 60 seconds; eviction policy `allkeys-lfu`; no jitter (short TTL) - **Validation**: Place order, verify count is updated in cache within 1 second of database commit; verify no negative inventory is displayed - [ ] **CACHE-ITEM-1.5 [API Response Cache Implementation]**: - **Description**: Cache aggregated API responses (dashboard, recommendations) for anonymous users. Key: `api:response:{endpointHash}:{queryParamsHash}`. - **Invalidation Strategy**: Cache-aside with short TTL; event-driven invalidation on underlying data change - **TTL and Eviction**: TTL 5 minutes; eviction policy `allkeys-lfu`; jittered TTL (±15 seconds) - **Validation**: Compare cached vs fresh API response; verify cache hit rate >90% for dashboard endpoint under load test ## Proposed Code Changes ### File: `src/main/java/com/example/cache/RedisConfig.java` ```diff +@Configuration +public class RedisConfig { + @Bean + public RedisConnectionFactory redisConnectionFactory() { + RedisClusterConfiguration clusterConfig = new RedisClusterConfiguration() + .clusterNode("redis-cluster-1.example.com", 6379) + .clusterNode("redis-cluster-2.example.com", 6379) + .clusterNode("redis-cluster-3.example.com", 6379); + clusterConfig.setPassword(RedisPassword.of("${REDIS_PASSWORD}")); + return new LettuceConnectionFactory(clusterConfig); + } + + @Bean + public RedisTemplate redisTemplate() { + RedisTemplate template = new RedisTemplate<>(); + template.setConnectionFactory(redisConnectionFactory()); + template.setKeySerializer(new StringRedisSerializer()); + template.setValueSerializer(new ByteArrayRedisSerializer()); + return template; + } + + @Bean + public CacheManager cacheManager(RedisConnectionFactory factory) { + RedisCacheConfiguration config = RedisCacheConfiguration.defaultCacheConfig() + .entryTtl(Duration.ofMinutes(60)) + .serializeValuesWith(SerializationPair.fromSerializer(new GenericJackson2JsonRedisSerializer())) + .disableCachingNullValues(); + return RedisCacheManager.builder(factory) + .cacheDefaults(config) + .build(); + } +} ``` ### File: `src/main/java/com/example/cache/UserProfileCacheService.java` ```diff +@Service +public class UserProfileCacheService { + private static final String KEY_PREFIX = "user:profile:"; + private static final Duration TTL = Duration.ofHours(24); + private static final Duration TTL_JITTER = Duration.ofSeconds(30); + + @Autowired + private RedisTemplate redisTemplate; + + @Autowired + private ObjectMapper objectMapper; + + public UserProfile getProfile(String userId) { + byte[] cached = redisTemplate.opsForValue().get(KEY_PREFIX + userId); + if (cached != null) { + return deserialize(cached); + } + UserProfile profile = loadFromDatabase(userId); + if (profile != null) { + setProfile(userId, profile); + } + return profile; + } + + public void setProfile(String userId, UserProfile profile) { + byte[] serialized = serialize(profile); + Duration ttlWithJitter = TTL.plus(Duration.ofMillis(ThreadLocalRandom.current().nextLong(-TTL_JITTER.toMillis(), TTL_JITTER.toMillis()))); + redisTemplate.opsForValue().set(KEY_PREFIX + userId, serialized, ttlWithJitter); + } + + @EventListener + public void handleProfileUpdate(UserProfileUpdatedEvent event) { + // Write-through invalidation + setProfile(event.getUserId(), event.getNewProfile()); + // Publish invalidation event for other services + redisTemplate.convertAndSend("cache-invalidation", "user:" + event.getUserId()); + } + + private UserProfile loadFromDatabase(String userId) { + // Database query implementation + } + + private byte[] serialize(UserProfile profile) { + // Protocol Buffers serialization + } + + private UserProfile deserialize(byte[] data) { + // Protocol Buffers deserialization + } +} ``` ### File: `src/main/java/com/example/cache/CacheStampedePreventionAspect.java` ```diff +@Aspect +@Component +public class CacheStampedePreventionAspect { + @Autowired + private RedisTemplate redisTemplate; + + @Around("@annotation(CacheStampedeProtected)") + public Object preventStampede(ProceedingJoinPoint joinPoint) throws Throwable { + // Implement mutex lock with Redis SETNX + String lockKey = "lock:" + joinPoint.getSignature().getName(); + Boolean acquired = redisTemplate.opsForValue().setIfAbsent(lockKey, "locked", Duration.ofSeconds(2)); + if (acquired != null && acquired) { + try { + return joinPoint.proceed(); + } finally { + redisTemplate.delete(lockKey); + } + } + // Wait and retry with exponential backoff + Thread.sleep(50); + return joinPoint.proceed(); + } +} ``` ## Commands ### Local Development ```bash # Run Redis locally for development docker run -d -p 6379:6379 redis:7-alpine # Run tests with local Redis mvn test -Dspring.profiles.active=local # Run integration tests with caching validation mvn verify -Dspring.profiles.active=test -Dskip.unit.tests=false # Load test with cache stampede scenario ./scripts/load-test.sh --scenario stampede --concurrency 100 --duration 60s # Benchmark cache hit rates ./scripts/benchmark-cache.sh --endpoint /api/products --warmup 1000 --requests 10000 # Check Redis cluster health redis-cli -h localhost -p 6379 cluster info # Monitor cache metrics ./scripts/monitor-cache.sh --metrics hits,misses,evictions,memory ``` ### CI Pipeline ```bash # In CI: run with ephemeral Redis container docker run -d --name redis-test -p 6379:6379 redis:7-alpine mvn test -Dspring.profiles.active=ci docker stop redis-test # CI validation of cache invalidation ./scripts/validate-invalidation.sh --endpoint /api/users/{id} --change-frequency 100ms # CI performance gates mvn gatling:test -Dgatling.simulationClass=CachePerformanceSimulation ``` ## Quality Assurance Task Checklist - [ ] All cache layers are documented with technology, configuration, and data flow - [ ] Invalidation strategies are defined for every cached data type (write-through for profiles, event-driven for catalog, cache-aside for sessions, write-behind for inventory) - [ ] TTL values are justified with data volatility analysis (24h for stable profiles, 1h for catalog, 5m for sessions, 60s for inventory) - [ ] Failure scenarios are handled with graceful degradation paths (fallback to database, circuit breakers, local cache as interim) - [ ] Monitoring and alerting covers hit rates, latency, memory, and eviction metrics - [ ] Cache key schema is documented with naming conventions and versioning (`user:profile:`, `product:catalog:`, `session:`, `inventory:count:`) - [ ] Performance benchmarks validate that caching meets target SLAs (p50 < 10ms, p99 < 50ms) - [ ] Security controls (TLS, authentication, network isolation) are applied to all cache endpoints

Estrategia de Arquitectura de Caché

Contenido del PromptGuardar

Uso

Referencias

Discusión