Infrastructure
Written for technical readers, and for anyone running or considering running their own PDS. For the conceptual side, see How it works.
- The request path covers Cloudflare, Caddy, TLS and the PDS itself.
- Keeping abuse out covers the controls in front of account creation and sensitive endpoints.
- Automation and operations covers scheduled jobs, updates, email and incident history.
The stack
Section titled “The stack”| Layer | What runs |
|---|---|
| Edge | Cloudflare, in Full SSL mode |
| Reverse proxy | Caddy, terminating TLS at the origin |
| Endpoint protection | PDS Gatekeeper, in front of sensitive XRPC endpoints |
| Application | The reference bluesky-social/pds, on localhost:3000 |
| Host | A single virtual machine on Microsoft Azure |
| Container updates | Watchtower |
| Scheduled jobs | Azure Automation runbooks, in PowerShell |
Nothing here is bespoke at the application layer. Tophhie Social runs the same reference PDS implementation Bluesky publishes, and the interesting decisions are all in what surrounds it.
Design priorities
Section titled “Design priorities”Three things shape the choices described in this section.
It is a personal-scale service, with 14 active repositories as of October 2025, but run to production expectations: 99.9% uptime in September 2025, and published post-incident reports when it fails.
Abuse resistance comes ahead of convenience. An open PDS with no invite codes is an attractive target, and has been attacked. See Keeping abuse out.
Destructive actions keep a human in the loop. Account deletion is reviewed and executed by a person, and automation handles the schedule rather than the judgement.
Open source
Section titled “Open source”The user-facing tooling around the PDS, meaning the dashboard, the migration tool and the sign-up tool, derives from the Witchcraft Systems PDS tooling. Tophhie Cloud open-sourced its own code in September 2025.
Sources
Section titled “Sources”This section is assembled from Tophhie Cloud’s published writing, primarily:

