Caching invalidation strategies

strategy requires explicit consistency boundaries, idempotent mutation propagation, and stampede mitigation. The following architecture uses a cache-aside pattern with hybrid invalidation, implemented in TypeScript with ioredis.

Step 1: Define Consistency Boundaries

Classify data by consistency requirements:

• Strong consistency: Financial balances, inventory counts, session state
• Eventual consistency: Product catalogs, user profiles, analytics aggregates Map these to invalidation tolerances. Strong data requires event-driven invalidation with mutex protection. Eventual data can tolerate TTL + versioned keys.

Step 2: Implement Cache-Aside with Versioning

Versioned keys prevent stale reads during deployments and enable selective invalidation. Instead of deleting keys, increment a version counter and append it to the cache key. Old versions naturally expire via TTL.

import Redis from 'ioredis';

const redis = new Redis({ 
  host: process.env.REDIS_HOST, 
  port: parseInt(process.env.REDIS_PORT || '6379'),
  maxRetriesPerRequest: 3,
  retryStrategy: (times) => Math.min(times * 50, 2000)
});

interface CacheConfig {
  ttlSeconds: number;
  versionKey: string;
  consistency: 'strong' | 'eventual';
}

export class CacheManager {
  constructor(private redis: Redis, private config: CacheConfig) {}

  async getVersion(): Promise<number> {
    const v = await this.redis.get(this.config.versionKey);
    return v ? parseInt(v, 10) : 0;
  }

  async invalidateVersion(): Promise<void> {
    await this.redis.incr(this.config.versionKey);
  }

  buildKey(entityId: string): string {
    return `${this.config.versionKey}:${entityId}:${this.config.consistency}`;
  }

  async get<T>(entityId: string): Promise<T | null> {
    const version = await this.getVersion();
    const key = `${this.buildKey(entityId)}:${version}`;
    const raw = await this.redis.get(key);
    return raw ? JSON.parse(raw) : null;
  }

  async set<T>(entityId: string, data: T): Promise<void> {
    const version = await this.getVersion();
    const key = `${this.buildKey(entityId)}:${version}`;
    await this.redis.set(key, JSON.stringify(data), 'EX', this.config.ttlSeconds);
  }
}

Step 3: Wire Event-Driven Invalidation

For strong consistency, publish invalidation events to a Redis Pub/Sub channel or message broker. Subscribers delete or version-bump affected keys. Use idempotent handlers to prevent duplicate processing.

import { EventEmitter } from 'events';

export class InvalidationBus extends EventEmitter {
  constructor(private redis: Redis, private channel: string) {
    super();
    this.subscribe();
  }

  private subscribe(): void {
    const subscriber = this.redis.duplicate();
    subscriber.subscribe(this.channel, () => {
      console.log(`[InvalidationBus] Subscribed to ${this.channel}`);
    });

    subscriber.on('message', async (ch, payload) => {
      if (ch !== this.channel) return;
      try {
        const { entityId, type } = JSON.parse(payload);
        this.emit('invalidate', { entityId, type });
      } catch (err) {
        console.error('[InvalidationBus] Parse error:', err);
      }
    });
  }

  async publish(entityId: string, type: 'UPDATE' | 'DELETE'): Promise<void> {
    await this.redis.publish(this.channel, JSON.stringify({ entityId, type }));
  }
}

Step 4: Mitigate Cache Stampedes

When a high-traffic key expires, concurrent requests trigger simultaneous database queries. Prevent this with probabilistic early expiration or a distributed mutex.

import { v4 as uuidv4 } from 'uuid';

export class StampedeGuard {
  constructor(private redis: Redis, private lockTtlMs: number = 3000) {}

  async acquireLock(key: string): Promise<string | null> {
    const lockKey = `lock:${key}`;
    const token = uuidv4();
    const acquired = await this.redis.set(lockKey, token, 'PX', this.lockTtlMs, 'NX');
    return acquired ? token : null;
  }

  async releaseLock(key: string, token: string): Promise<void> {
    const lockKey = `lock:${key}`;
    const current = await this.redis.get(lockKey);
    if (current === token) {
      await this.redis.del(lockKey);
    }
  }
}

Architecture Decisions & Rationale

• Cache-aside over write-through: Write-through synchronously blocks application threads until cache and DB commit. Cache-aside defers cache population to read paths, reducing write latency and simplifying error handling.
• Versioning over direct deletion: Deleting keys forces immediate rebuilds under load. Versioning allows old keys to expire naturally while new reads fetch fresh data, eliminating stampede pressure.
• Pub/Sub for invalidation: Low-latency, zero-persistence broadcast. Acceptable for invalidation because duplicate or missed messages are handled by TTL fallback and idempotent version increments.
• Distributed mutex for stampedes: Redis SET NX PX provides lightweight mutual exclusion without external coordination services. Fallback to direct DB query if lock acquisition fails after timeout.

Pitfall Guide

1. Blind TTL extension on read: Extending TTL every time a key is accessed creates "hot" keys that never expire, causing stale data to persist indefinitely. Best practice: Only extend TTL on explicit writes or use sliding expiration with a hard maximum lifetime.

2. Cross-entity invalidation gaps: Updating a parent entity without invalidating cached child aggregates or denormalized views. Best practice: Map entity relationships explicitly and trigger batch invalidation events for all dependent cache keys.

3. Thundering herd on expiration: Thousands of requests hitting an expired key simultaneously. Best practice: Implement probabilistic early expiration (e.g., 10% chance to refresh 20% before TTL expiry) or distributed mutex locks for high-traffic keys.

4. Race conditions in write-through: Application writes to DB, cache updates, but a concurrent read fetches stale cache before invalidation propagates. Best practice: Use cache-aside for write-heavy paths, or enforce sequential consistency via distributed locks around critical sections.

5. Ignoring cache topology: Redis Sentinel or Cluster modes introduce replication lag and partition tolerance. Invalidation messages may arrive on nodes that haven't replicated the mutation yet. Best practice: Route invalidation through a single authoritative node or use consistent hashing with version vectors.

6. Unbounded invalidation queues: Message brokers backing up during traffic spikes, causing delayed invalidation and memory pressure. Best practice: Implement backpressure, discard stale invalidation events older than a configurable threshold, and monitor queue depth.

7. No cache health observability: Teams monitor hit rate but ignore stale read ratio, invalidation latency, and stampede frequency. Best practice: Instrument cache managers with metrics for cache_stale_ratio, invalidation_propagation_ms, and stampede_prevented_count. Alert on stale ratio > 2% or invalidation latency > 500ms.

Production Bundle

Action Checklist

• Classify data by consistency requirements: Map each cached entity to strong or eventual consistency bounds before deployment.
• Implement versioned cache keys: Append incrementing version counters to keys instead of deleting them on mutation.
• Wire idempotent invalidation handlers: Ensure duplicate Pub/Sub messages or queue retries do not cause side effects or double-deletes.
• Add stampede mitigation: Deploy distributed mutex locks or probabilistic early expiration for keys exceeding 500 RPS.
• Configure TTL fallbacks: Set conservative TTLs (30–300s) on all keys to recover from silent invalidation failures.
• Instrument cache metrics: Track hit rate, stale read ratio, invalidation latency, and DB query amplification per cache tier.
• Test invalidation paths: Run chaos experiments that kill cache nodes, delay Pub/Sub, and inject duplicate events to verify recovery.

Decision Matrix

Scenario	Recommended Approach	Why	Cost Impact
Read-heavy catalog with infrequent updates	TTL-Only + Versioning	Low write volume makes event-driven overhead unnecessary; versioning handles deployments safely	Low infra cost, moderate memory usage
Inventory/financial data requiring sub-100ms consistency	Event-Driven + Mutex	Strong consistency demands immediate invalidation; mutex prevents stampede-induced DB thrashing	Higher network overhead, reduced DB load during bursts
Multi-region deployment with cross-node replication lag	Hybrid + Version Vectors	Pub/Sub latency varies across regions; version vectors ensure stale reads are detected and refreshed	Moderate infra cost, requires version sync service
High-churn session/cache with strict memory limits	TTL-Only with aggressive eviction	Memory constraints favor expiration over explicit invalidation; versioning adds key bloat	Low operational overhead, requires careful TTL tuning

Configuration Template

// cache-config.ts
import { Redis } from 'ioredis';
import { CacheManager } from './cache-manager';
import { InvalidationBus } from './invalidation-bus';
import { StampedeGuard } from './stampede-guard';

const redis = new Redis({
  host: process.env.REDIS_HOST || '127.0.0.1',
  port: parseInt(process.env.REDIS_PORT || '6379'),
  password: process.env.REDIS_PASSWORD || undefined,
  maxRetriesPerRequest: 3,
  retryStrategy: (times) => Math.min(times * 100, 2000),
  enableReadyCheck: true,
  connectTimeout: 5000
});

export const cacheManager = new CacheManager(redis, {
  ttlSeconds: 120,
  versionKey: 'app:version:products',
  consistency: 'eventual'
});

export const invalidationBus = new InvalidationBus(redis, 'invalidation:products');

export const stampedeGuard = new StampedeGuard(redis, 2500);

// Graceful shutdown
process.on('SIGTERM', async () => {
  await redis.quit();
  process.exit(0);
});

Quick Start Guide

1. Initialize Redis: Run docker run -d -p 6379:6379 redis:7-alpine to start a local Redis instance.
2. Install dependencies: Execute npm i ioredis uuid in your project directory.
3. Copy configuration: Paste the Configuration Template into cache-config.ts and ensure environment variables match your Redis endpoint.
4. Run integration test: Create a test script that calls cacheManager.set(), publishes an invalidation event, verifies cacheManager.get() returns updated data, and confirms version increment. Execute with npx ts-node test-cache.ts.
5. Verify metrics: Check Redis keyspace hits/misses via redis-cli INFO stats and confirm invalidation latency stays under 200ms under 1k RPS load testing with autocannon.