Caching Strategies Guide
Caching is layers: the browser, the CDN edge, an application cache like Redis, and the database's own memory. Each layer trades freshness for speed at a different price.
Decide a staleness budget per data type before touching code — marketing pages can lag minutes, a shopping cart cannot lag at all. The budget tells you which layer is allowed to cache what.
HTTP and CDN caching
Static assets with hashed filenames get Cache-Control: public, max-age=31536000, immutable. Dynamic HTML and APIs use short s-maxage values plus stale-while-revalidate so the CDN serves instantly and refreshes in the background. Cloudflare caches static files by default and caches HTML or API responses when you opt in via cache rules; purge by URL is universal, while tag-based purge exists on Fastly and Cloudflare's enterprise tier. Send ETags so clients revalidate cheaply.
Application caching with Redis and KV
The workhorse pattern is cache-aside: read the cache, miss, load from the database, write back with a TTL. Redis — Upstash for serverless, ElastiCache on AWS, Memorystore on GCP — gives strong consistency and rich structures for counters and leaderboards. Cloudflare KV is different: reads are fast at every edge, but writes propagate globally over tens of seconds, which suits config and feature flags, not anything transactional.
Invalidation that actually works
Prefer boring TTLs over clever purging; most data tolerates thirty seconds of staleness. Where precision matters, use versioned keys — bump user:42:v7 to v8 on write and old entries expire naturally. Protect hot keys from stampedes with a single-flight lock or by serving stale while one worker refreshes. Never cache authenticated responses at a shared CDN unless they are keyed per user.
What not to cache
Skip caching for money movements, inventory counts near zero, permission checks and anything feeding a write decision. A stale read that triggers a wrong write costs more than every millisecond caching ever saved. When in doubt, measure first: cache your three slowest queries, not everything.
Frequently asked questions
Should I use Redis or Cloudflare KV?
Redis for consistency, counters and read-after-write needs. KV for read-heavy, rarely-written data like config, where seconds of propagation delay are harmless.
How long should TTLs be?
As long as your staleness budget allows: minutes for catalogs, seconds for dashboards, zero for balances. Start short and lengthen with evidence.
What is a cache stampede?
A hot key expires and hundreds of requests hit the database at once. Prevent it with single-flight locking or stale-while-revalidate.
Should I cache at the CDN or in the app?
Both, for different data: the CDN for shared public responses, the app cache for per-user or composed data. The layers complement rather than compete.