Fork me on GitHub

After a recent update to one of my CentOS servers that hosts some django applications, I started getting 502’s with the following error message:

2013/05/05 10:06:04 [alert] 1771#0: *22 socket() failed (97: Address family not supported by protocol) while connecting to upstream, client: X.X.X.X, server:, request: "GET /bid/ HTTP/1.1", upstream: "http://[::1]:8001/bid/", host: ""

To make matters worse, I was only seeing this error message intermittently. I could refresh the page 10 times and only see this error 5/10 times.

Everything I found online said something about commenting out a line like this one in your nginx config:

listen   [::]:80 default ipv6only=on;

However, I didn’t have such a line in any of my configs. Then I noticed the following snippet from the error message above:

upstream: "http://[::1]:8001/bid/"

and realized that nginx (or, more likely, the system routing code) was resolving localhost to ::1 (which is localhost in ipv6)! I considered editing my /etc/hosts file to remove the ::1 alias from localhost but then settled on a cleaner solution: replacing all instances of localhost in my nginx config files with

The moral of the story: localhost might not mean what you think it means! Don’t use localhost unless you really support both ipv4 and ipv6. Instead, use and ::1.

blog comments powered by Disqus