Kingsfordweg 151, 1043GR Amsterdam, The Netherlands ● 24/7 support — support@novogara.com
Novogara BV
Knowledge base

Caching and CDNs

A CDN is often sold as the fix for a slow site. Frequently the fix is three response headers you already control.

Headers first

Nothing caches anything unless you say so.

  • Fingerprinted assets — app.a1b2c3.css — get Cache-Control: public, max-age=31536000, immutable. A year, because a new build gets a new name.
  • Images and fonts that rarely change: weeks to months.
  • HTML: short, or no-cache with an ETag so clients revalidate cheaply.
  • Anything personal: Cache-Control: private, no-store. Caching a logged-in page publicly serves one user's data to another.

Then check what you actually send:

curl -sI https://example.com/app.css | grep -i cache

Your own edge with nginx

Before buying anything, cache on the machine you already rent:

proxy_cache_path /var/cache/nginx levels=1:2 keys_zone=edge:100m max_size=20g inactive=7d;

and in the location: proxy_cache edge; with proxy_cache_use_stale updating error timeout; so a slow backend does not mean a slow site. Add proxy_cache_lock on; to stop a thundering herd when a popular object expires.

On a machine with plenty of memory and an unmetered port, this handles very large volumes and costs nothing per gigabyte.

When a CDN genuinely helps

  • Your users are far away. Physics: a request from Sydney to Amsterdam is about 280 ms round trip no matter how fast your server is.
  • Very spiky traffic you do not want to provision for.
  • You want absorption of attacks at the application layer, beyond the volumetric DDoS filtering we already provide.

When it does not

  • Your users are European. From Amsterdam you are within 10–35 ms of most of them — see why Amsterdam. A CDN adds a hop and saves little.
  • Your content is dynamic and personal. There is nothing to cache.
  • Your site is slow because of the database. A CDN will not touch that; see diagnosing a slow server.

Egress is the quiet argument

CDNs are usually priced per gigabyte delivered. On an unmetered port, delivering from your own machine has a fixed monthly cost no matter how busy the month is. For high-volume delivery that difference is frequently the largest number in the comparison — see streaming and high-bandwidth delivery.

Measure before and after

Take a real figure first — time to first byte and full load, from where your users are. Then change one thing. “It feels faster” is how people end up paying monthly for a layer that made no difference.

Still stuck? Mail support@novogara.com — an engineer answers, at any hour. Back to the knowledge base