Advanced

Cache Invalidation

Caching Patterns

TTL, event-driven, and manual invalidation to manage cache freshness

Core Idea

Control staleness by expiring, versioning, or actively invalidating cached data. Use TTL with jitter, event-driven invalidation, and versioned keys to balance freshness and performance.

When to Use

When correctness matters with cached data, or you operate multiple cache layers and need predictable freshness.

Recognition Cues
Indicators that this pattern might be the right solution
  • Users see stale data after updates
  • Cache stampede on popular keys
  • Manual flushes frequently needed
  • Mismatched data across layers (browser/CDN/app cache)

Pattern Variants & Approaches

Overview
Write-through or write-behind updates publish invalidation events; workers evict or update keys across cache tiers.

Overview Architecture

WriteReadPublishConsumeEvict/Update⚙️Application💾Database⚡Cache📬Event Bus⚙️Invalidation Worker

When to Use This Variant

  • Stale data anomalies
  • Multi-layer caches (browser/CDN/app)
  • Need precise freshness windows

Use Case

Feeds, product catalogs, and dashboards requiring fresh-but-fast reads.

Advantages

  • Predictable freshness
  • Reduced stampedes
  • Clear ownership of keys

Implementation Example

# Invalidation flow (pseudocode)
db.update(row)
eventBus.publish({ type: 'invalidate', key })
worker.on('invalidate', key => cache.delete(key))
Tradeoffs

Pros

  • Predictable freshness and fewer anomalies
  • Reduced origin load and stampedes
  • Better multi-layer cache coherence

Cons

  • More moving parts and state to track
  • Eventual consistency windows
  • Operational complexity and tooling needs
Common Pitfalls
  • No TTLs or very long TTLs by default
  • Inconsistent key naming/namespacing
  • Not using negative caching for misses
  • Event-driven invalidation lost or unordered
  • No single-flight leading to thundering herd
Design Considerations
  • Set TTLs with jitter; add soft TTL and SWR
  • Version keys with schema/content hash
  • Event-driven invalidation via CDC/pub-sub
  • Write fences to prevent stale reads post-write
  • Request coalescing/single-flight for hot keys
Real-World Examples
Facebook

TAO graph cache invalidation at scale

Billions of entities
Reddit

Event-driven invalidation of hot pages

Hundreds of millions of users
Etsy

Versioned cache keys for releases

Large commerce catalogs
Complexity Analysis
Scalability

Horizontal - Distributed caches

Implementation Complexity

High - Hardest problem in CS

Cost

Medium - Infra + engineering time

Related Patterns
Caching StrategiesMessage QueuesCQRSEvent Sourcing