Zero downtime deployments

[stancl]

Can be handled either with multiple upstreams in nginx (where putting an app down via php artisan down — returning 5xx — makes nginx use the other upstream) or by atomically swapping a symlink — blue/green strategy.

Some considerations:

• Need some synchronization, can't have deployments running concurrently*
• Migrations need to be backwards-compatible (i.e. DROP only in the n+2nd deployment after n+1 serves all requests)
• Shared assets — file-based cache, SQLite databases, user data storage etc — need to be either symlinked or just use absolute paths to some location outside these deployed versions, perhaps something in /var
• Could use something like /srv/site-foo symlinked to /srv/site-foo-{n}

* Technically not unique to zero downtime deployments, good practice overall, but could be riskier here. Implementing this could be as simple as flock -w 100 /tmp/deployment_{sitename} -c "bash deploy.sh"