Tukuk-OSThe One Idea That Matters
A CDN stores responses at edge locations near users. The hard part is not storing — it's knowing how long that stored copy may still be served. That decision is a TTL (time-to-live) and a cache key.
- Cache key = what identifies "the same" response. Usually the URL, but also the parts you vary on (host,
Accept, cookies, a version header). Get this wrong and you serve the logged-in page to anonymous users, or you never reuse the cache at all.
- TTL = how long the edge may serve without checking origin. Short TTL = fresher but more origin load. Long TTL = cheaper but staler.
Setting TTLs by Volatility
Cache the things that rarely change, longest:
| Content | Typical TTL |
|---|---|
| Versioned static assets (app.a1b2c3.js) | Weeks–a year (they're immutable by name) |
| Images / fonts | Days–weeks |
| HTML shell | Minutes, or no-cache + revalidate |
| API responses | Seconds–minutes, or don't cache |
Immutable, fingerprinted filenames let you cache forever: you never invalidate, you just ship a new name.
Serving Stale on Purpose (SWR)
stale-while-revalidate lets the edge serve the expired copy immediately while it fetches a fresh one in the background. Users get instant (slightly old) content and the cache quietly updates. Perfect for feeds, product pages, blog lists — where one-minute-stale is invisible but sub-second latency is felt. stale-if-error does the same when the origin is down, which turns an outage into "looks a little dated" instead of "502 everywhere".
``http`
Cache-Control: public, max-age=60, stale-while-revalidate=300
Freshness for 60s, then serve stale for up to 300s while revalidating.
When You Must Purge (and When Not To)
Purge (invalidate) is a blunt, per-object operation you reserve for real need: a price changed, a security page updated, content deleted. Purging everything on every deploy is an anti-pattern — it turns your CDN into a stampede against origin (cache miss = every edge races the source at once, a thundering herd). Prefer:
- Versioned names for cache-busting (never purge).
- Short TTL + SWR for pages that just need to be roughly current.
- Targeted purge for the few URLs that genuinely must drop now.
Debugging Checklist
- Why is my user seeing old content? Check the cache key first — are you varying on something you forgot (user, device, cookie)? Then the TTL.
- Why is origin getting hammered? Cache key too specific, TTL too short, or you're purging globally.
- Is anything actually cacheable? If responses set Cache-Control: private
or carry aSet-Cookie`, edges won't store them.
A CDN is a distributed set of promises about freshness. Keep the cache key honest, match TTL to volatility, serve stale deliberately, and purge only when a human would notice the difference.