Caching Strategy

Caching Strategy

Purpose

Add caching deliberately — with a clear read/write/invalidate flow, stampede protection, and an invalidation plan — so it reduces load without serving stale or inconsistent data or collapsing under a hot-key herd.

Universal — cache-aside flow, TTL+jitter, stampede prevention, and invalidation strategy are caching principles independent of the cache store; Redis is the default implementation.

Procedure

1. Optimize the query FIRST, cache second

   • Caching a slow query hides the problem and adds staleness risk
   • Run query-optimization before adding a cache layer

2. Use cache-aside (lazy loading) as the default pattern

   • Read: check cache → miss → read DB → populate cache → return
   • Write: write DB → invalidate (delete) the cache key (don't write-through unless justified)
   • Delete-on-write > update-on-write: avoids cache/DB races

3. Set TTL with jitter

   • Every cached key gets a TTL (no infinite caches without an invalidation plan)
   • Add random jitter to TTLs so keys don't all expire simultaneously (a synchronized expiry = mass stampede)

3b. Bound the cache: memory budget + eviction policy

• Set maxmemory and choose an eviction policy deliberately (allkeys-lru for general read-through caches; volatile-lru if you mix persistent state into the same instance — but ideally don't)
• Without a bound, a runaway key generator (per-user, per-query-fingerprint) eats memory until OOM
• Cache-key cardinality: unbounded distinct keys = unbounded memory; cap or hash high-cardinality identifiers

4. Prevent cache stampede on hot keys

   • Single-flight lock: first request acquires an atomic compare-and-set lock, recomputes, others wait/serve-stale
   • Probabilistic early refresh: recompute slightly before expiry with rising probability
   • Both must be executed atomically (store-specific mechanism in Implementation)
   • Without this, a popular key expiring under load → thundering herd hammers the DB
   • Negative caching: cache "this key doesn't exist" (a sentinel value, short TTL) for queries that miss — otherwise the same non-existent key hits the DB on every request (silent thundering herd from 404s)

5. Plan invalidation explicitly — the hard part

   • Know exactly which writes invalidate which keys
   • Use key naming conventions (user:{id}:profile) so invalidation is targeted
   • Tag-based / versioned keys for "invalidate everything related to X"

5b. Cache is an optimization, not a source of truth

• The app MUST keep working with an empty / unavailable cache — degraded latency, not broken behavior
• Never store authoritative state only in the cache (auth tokens, balances). On Redis restart you lose it
• Use resilience-patterns circuit breaker around the cache client so a cache outage doesn't cascade

6. Validate (validation loop)
   • Load-test with a hot key expiring under concurrency → verify no DB spike (stampede prevented)
   • After a write, verify the cache returns fresh data (invalidation works)
   • If stale data served → invalidation gap; fix the write→invalidate wiring

Anti-patterns

Severity tiers

Completion Criteria

• Underlying query optimized before caching
• Cache-aside with delete-on-write applied
• Every key has a TTL + jitter
• Hot keys have stampede protection (verified under load)
• Invalidation mapping documented (which write → which key)

Output

• Cache layer code: cache-aside helpers + invalidation hooks
• Invalidation map: docs/cache-invalidation.md — write → invalidated keys
• Commit format: perf(cache): add cache-aside for <query> / fix(cache): single-flight lock on <hot key>

Implementation

TypeScript + Redis (default)

• Cache-aside helper around Redis GET/SETEX/DEL
• Stampede: single-flight via SET key val NX PX ttl lock, or ioredis + Lua EVAL for atomicity
• Probabilistic refresh: store (value, computed_at, ttl) and recompute when now - computed_at > ttl * random_threshold
• Supabase: pair with Postgres; Redis via Upstash/managed

Other stacks

• Python: redis-py + aiocache; same cache-aside + Lua patterns
• Go: go-redis + singleflight package (stdlib-adjacent) for stampede
• Universal: cache-aside, TTL+jitter, and stampede prevention are store-agnostic; Memcached works for simple cases (no Lua → use add-based locks)

Related skills

• query-optimization — cache only after the query itself is optimized
• resilience-patterns — cache as a fallback when a dependency is down
• transaction-management — invalidate cache after the write commits, not before

Reference

• Key insight encoded: A popular key expiring under concurrency triggers a stampede (thundering herd); prevent it with a single-flight mutex lock or probabilistic early refresh, both executed atomically via Lua. Delete-on-write (not update) avoids cache/DB races.