Narayan Newton

I'm going to be honest, I'm kind of obsessed with caching. Since becoming involved in web performance, it has pretty much taken over my life. Yes there is the mysql tuning, query profiling, solr deployments....but all in all 90% of my day is looking at hit rates of one form or another. Due to this, HTTP caching for drupal.org has been a mild obsession for me. This week my wish came very close to true.

We have finally deployed HTTP level caching on drupal.org. To give some background, we have run Squid as a reverse proxy for quite awhile now. It has worked well for us in caching static files and acting as a connection pool. However, it has never done well caching actual drupal generated pages. Every anonymous user gets a SESSION cookie and that prevents any sort of caching.

Recently David Strauss and Damien backported the D7 lazy session initialization and HTTP caching headers to D6 for drupal.org. This was the first glimmer of hope that we might actually have "Sane Caching" soon. With this support, we can set the HTTP caching headers to Vary on the Cookie header and since anonymous users don't get Cookies, their pages can be cached at the reverse proxy and in fact any caching proxies in between the server and the client.

The only road block to a 'Caching Utopia' was that we had to strip out the google analytics cookies. These arn't used on the server side, only the client has any use for them, but they do mess up our master plan. Stripping these cookies is difficult in squid and with them all of our great patches to drupal were useless. This is where Varnish enters the story. I've been watching Varnish with interest for some time. Unlike squid, Varnish was built to be a reverse caching proxy (while squid does it on the side to pay for its various habits). I prefer the design of varnish and how simple it is. I could see from the beginning that it had the potential to be far faster and far lighter than squid, but it was unstable. Until recently. After successfully deploying Varnish2 on our solr cluster I decided to give it a try on the web nodes and use it to strip the analytics cookies off. The results have been impressive.

The above is the CPU usage on www3.drupal.org. As you can plainly see, the difference between Squid/No Caching and Varnish/Anonymous Caching was significant. While we are still resolving some issues after the switch, it has been a large success on the whole and has lowered the load on our web nodes across the board.

Varnish (http://varnish.projects.linpro.no/)
Squid (http://www.squid-cache.org/)
D.O. Issue (http://drupal.org/node/466444)
Patch BZR Tree/Distro (http://pressflow.org)