DeanKelly751 Caching Strategy
Purpose
Step-by-step procedure for choosing and implementing effective caching.
Workflow
Step 1: Identify What to Cache
Good cache candidates:
- Expensive computations with stable inputs
- Frequently read, infrequently written data
- External API responses with acceptable staleness
- Database queries with predictable access patterns
Poor cache candidates:
- Highly personalized data (low hit rate)
- Data that must always be current
- Write-heavy data (invalidation overhead)
Step 2: Choose Cache Placement
| Layer | Latency | Scope | Example |
|---|---|---|---|
| In-process | ~ns | Single instance | HashMap, LRU cache |
| Distributed | ~ms | All instances | Redis, Memcached |
| CDN | ~ms | Global | Static assets, API responses |
| Browser | 0 | Single user | HTTP cache headers |
Step 3: Choose Strategy
| Pattern | How it works | Best for |
|---|---|---|
| Cache-aside | App checks cache, fills on miss | General purpose |
| Read-through | Cache fetches on miss | Simplify app code |
| Write-through | Write to cache and DB together | Strong consistency needs |
| Write-behind | Write to cache, async to DB | Write-heavy workloads |
Step 4: Design Invalidation
The two hard problems: cache invalidation and naming things.
| Approach | Consistency | Complexity |
|---|---|---|
| TTL-based | Eventual (bounded staleness) | Low |
| Event-based | Near-real-time | Medium |
| Write-through | Strong | Medium |
| Manual purge | On-demand | Low (but error-prone) |
Step 5: Prevent Stampedes
When a popular cache entry expires, many requests hit the backend simultaneously.
Mitigations:
- Lock/single-flight: Only one request refreshes, others wait
- Early refresh: Refresh before TTL expires (probabilistic early expiration)
- Stale-while-revalidate: Serve stale data while refreshing in background
Checklist
- Cache hit rate monitored (target: >80% for general caches)
- TTL set with acceptable staleness tolerance
- Cache stampede prevention for hot keys
- Graceful degradation when cache is unavailable
- Cache size bounded (eviction policy: LRU, LFU)
- Serialization format chosen (avoid language-specific formats)
Related Skills
performance/knowledge-caching-patterns— deeper caching theoryperformance/knowledge-optimization-principles— measure before cachingbackend/knowledge-concurrency— cache access under concurrency