WordPress now runs on PicoPaaS: what it took

We restored my old personal blog onto PicoPaaS end to end: a non-root nginx and PHP-FPM image, persistent volumes, presigned uploads and a one-time restore. Here is what it took, how patching works, and what is next.

Share

A while ago I wrote about whether WordPress could run on PicoPaaS at all. The answer then was "probably, with some platform work". We just did the work, and I restored my old personal blog onto a throwaway app end to end. Here is what it took, how the image stays patched, and what is still missing.

What it took

A different image. The official WordPress image runs Apache as root. PicoPaaS pods are non-root with every Linux capability dropped, so that was a non-starter. We built our own: nginx plus PHP-FPM, running as uid 10001 on port 8080, writing only to the WordPress root and /tmp. It is pinned to PHP 8.5.12 because of the critical PHP remote code execution bugs (CVE-2026-109618 and CVE-2026-109619) fixed on 21 October. Running an unpatched PHP on a public platform was not an option.

A locked-down first boot. An uninstalled WordPress lets whoever shows up first finish the web installer and become admin. So the installer pages always return 403, and the entrypoint runs wp core install itself. The admin password is generated inside the container and printed once in the app logs. The eight auth keys and salts are required secrets: if one is missing the container refuses to start, because a random fallback would log everyone out on every restart. The MySQL connection uses TLS, and wp-login.php is rate limited.

Persistent volumes. WordPress wants a writable filesystem that survives restarts, and PicoPaaS had none. So volumes are now a platform feature: one ReadWriteOnce block volume per app, mounted where you ask, single replica, Recreate deploys, grow-only, $0.10/GiB/month. Our security review caught a nice one: a mount path with a doubled leading slash skipped the deny-list for paths like /etc while the container runtime resolved it to the denied directory. It is fixed, with tests. The review also turned up a capacity-count race and a spec re-apply bug after an admin grows a volume. Both are fixed too.

Getting a bundle in. Migrating means moving a database dump and a wp-content tarball, and apps had no way to put objects into their own bucket. We added presigned uploads, so now it is one command, with no bucket credentials on your laptop:

picopaas storage upload myapp ./db.sql.gz --key restore/db.sql.gz

Then WP_RESTORE=1 does a one-time restore on boot: it drops existing tables, streams the dump into MySQL over TLS, replaces wp-content, runs wp search-replace for the new URL, writes a marker, and never runs again. The site answers 503 while it works, and a failure leaves it at 503 with the reason in the logs.

The rehearsal

I restored a real 2015-era blog onto a throwaway app with a 1 GiB volume. Some things we learned:

  • The database dump shrank from 54.6 MB to 0.3 MB once a security-audit plugin's log tables had their rows dropped. Log-style data was the overwhelming majority of it and it was not the site.
  • One old plugin died on PHP 8 because it used create_function(), which no longer exists. WP_RESTORE_DEACTIVATE switches off the offenders after the restore.
  • A restart in the middle of a background auto-update left WordPress stuck in maintenance mode for visitors. Background updaters are now off by default in our image, and the entrypoint clears a stale .maintenance at boot.

The numbers: the restore took about two minutes after the redeploy. Idle, the pod sits around 2m CPU and 20 MiB. The peak I saw was 92m and 104 MiB. Public checks passed, the installer stayed blocked, a restart skipped the restore, and the uploads were still there. I deleted the bundle from the bucket afterwards.

Do we still need image builds?

Yes, and it is worth being clear about why. The image is the runtime: PHP-FPM, nginx, the extensions, WP-CLI. It is not just WordPress. With a volume, WordPress core, plugins and themes live on the volume after the first boot and you update them from wp-admin. Background auto-updates are off by default in our image (WORDPRESS_AUTOMATIC_UPDATES=true turns them back on). But PHP and OS security patches still only arrive through a new image.

The examples workflow rebuilds the image every Monday, pinned to the php:8.5.x-fpm-alpine tag. Alpine and OpenSSL fixes that Docker Hub republishes under that tag flow in on their own. PHP minor and patch versions, the seeded WordPress version and WP-CLI are bumped by hand in the Dockerfile. Apps pin an image tag and move forward with a redeploy, which with a volume means about a minute of downtime.

Tag pinning is a trade-off. A digest pin is reproducible, but it blocks those free base-image fixes and needs a manual bump for each one. We chose the tag and accept a little drift in exchange for patches arriving without anyone remembering.

What is next

Before I call WordPress first-class, and before the real migration, this is the list:

  • Backups. restic or kopia to an encrypted bucket, DigitalOcean's managed-database daily backups as the safety net, and optional volume snapshots. Right now the volume is not backed up.
  • One-off jobs and exec for imports and wp-cli, plus scheduled jobs per app.
  • A mail path. Outbound SMTP is blocked, so that means HTTPS-API mailers or a relay.
  • A custom domain for the migrated blog.
  • Boot-time admin actions such as a password reset or removing a plugin, which are in progress.
  • Volume usage in the health views.
  • Automated patch bumps for the Dockerfile, for example Renovate or Dependabot.

Then the production migration. I will write that one up too.