austin adams

Using git to Deploy a Hugo Blog... Atomically!

August 14, 2016

For years, I’ve been meaning to set up a blog, but characteristically I couldn’t muster the effort to even try. Yesterday, however, I found Hugo, a simple but useful static site generator, and finally decided to stop being lazy.

Despite the benefits (and despite writing a surprisingly similar tool myself ~4 years ago), I’d never felt drawn to static site generation. Often, I got stuck on how to (1) copy my changed content over to the webserver and (2) how to do it atomically. That is, when I update my blog, how do I easily delete the old files and copy over the new ones without any time period, however brief, during which some of the files are missing?

(For example, simply rm -ring /my/docroot/blog/* and then copying in the new files won’t work because between the time you delete some page X and copy it over, requests to X will 404.)

My answer, of course, is git combined with a kludgey atrocity some elegant git hooks!

The Process

  1. Locally, I change the source files for the site, which live in a git repository
  2. I commit my changes and push them over SSH to a bare git repository on the server
  3. On the server, the git post-receive hook runs, which:
    1. Checks out the new HEAD in a temporary directory
    2. cds into the temporary directory and runs Hugo with another new directory as the destination
    3. Replace the symlink from /my/docroot/blog to the old destination with a new symlink pointing to the new destination. Using the rename(2) syscall via mv makes this atomic.

Because I host cgit and this blog on the same server, I reused the existing public, bare repository as the remote repo which has the hook. And because I like to keep my docroot tidy, my hook creates all of the directories and links mentioned above in a separate directory, /var/www/blog/.

Here’s my hooks/post-receive:

set -e

# Store the built site snapshots outside of the docroot
# Use the SHA-1 of the current commit to name the new
# destination directory
destdir="$destroot/$(git rev-parse HEAD)"
olddestdir="$(readlink -e "$destlink" || true)"
worktree="$(mktemp -d)"

# Create a temporary working tree and delete it
# after building the site
GIT_WORK_TREE="$worktree" git checkout -f
pushd "$worktree"
    hugo --destination="$destdir"
rm -rvf "$worktree"

# Point to the new version of the site (atomically!)
# and delete the old version of the site
ln -svnrf "$destdir" "$destlinktmp"
# Use rename(2), which is atomic
mv -Tv "$destlinktmp" "$destlink"
rm -rvf "$olddestdir"

Then I simply point my HTTP server to the symlink, /var/www/blog/HEAD, like (in the case of nginx):

location /blog {
    alias /var/www/blog/HEAD/;

Tada! Now I can just git push to deploy changes to my blog.

A Warning on ln -f versus mv

My hook above creates a temporary link and uses mv to replace the old link with it. Why not use ln -f?

Well, as Tom Moertel pointed out back in 2005, ln -f ain’t atomic, at least with GNU coreutils on the Linux kernel. The following strace output for ln -s /x/y/z foo confirms this:

symlink("/x/y/z", "foo") = -1 EEXIST (File exists)
unlink("foo")            = 0
symlink("/x/y/z", "foo") = 0

So instead, I (and Tom) would suggest creating a temporary symlink and then moving it into place using mv, which uses rename(2), an atomic operation according to the manpage. strace output from mv -T foo.tmp foo:

rename("foo.tmp", "foo") = 0