Avoiding a broken site during deployment.
THE PROBLEM
A deployment that copies files one at a time leaves the site in a mixed state. A visitor arriving mid-deployment sees errors.
THE SIMPLE APPROACH
A Git pull is fast and mostly atomic in practice. For small sites this is usually sufficient.
THE BETTER APPROACH
Deploy into a new directory, then switch a symlink to point at it. The switch is instantaneous, and rolling back means switching the symlink back.
This requires the web server's document root to point at a symlink, which is configurable.
SHARED DIRECTORIES
Uploads, storage and configuration must persist across deployments. Keep them outside the release directories and symlink them in.
DATABASE MIGRATIONS
The hard part. A migration that removes a column breaks the old code still running during the switch.
Write migrations that are compatible with both old and new code where possible: add before removing, and remove in a later deployment.
MAINTENANCE MODE
For deployments that cannot be made seamless, a brief maintenance page is more honest than errors.
