Atomic website update with Git

Question

I'd like to update a live website in an atomic way. If we've made several changes to pages, images, and javascript, all of the changes should appear simultaneously with no downtime, inconsistency, or 404s.

If I simply copy the changed files into the website directory, then at least a few requests will see an inconsistent site. If I copy the new site to a separate directory, then rename the old directory to take it out of production and rename the new directory to make it live, then for one brief moment we'll get 404s.

I really have two questions:

How can I update a directory of files in an atomic way?

How can I coordinate this with Git? We'd like to deploy using a git pull (or possibly push). The path of the site within the git repo is different from the path on the server, so a bit of moving/renaming needs to happen, either using git commands or OS commands.

Answer 1

You can do what you're imagining using symlinks. Moving one symlink over another is an atomic operation, so you should be able to avoid any 404 errors.

Your hook would deploy your site to a directory, perhaps named after the commit hash. Then it would make a symlink to that, perhaps called staging. Then it would move that symlink over the production symlink.

hash=`git rev-parse HEAD`
git checkout-index -a -f --prefix=/srv/www/$hash/
ln -s /srv/www/$hash /srv/www/staging
mv -T /srv/www/staging /srv/www/production

The -T argument is the short form for --no-target-directory, which is a part of GNU mv and prevents it from moving the source into the destination. Read more about target directories.