An agency with forty customer sites, a company with a public site, a partner portal and an internal tool, a product family with a site per product: FFS serves all of them from one installation — without letting one site see, slow down or break another.

One server, one process per site

                       :443  :80
                          │
              ┌───────────▼────────────┐
              │      FFS front-end     │  TLS, routing, guard,
              │                        │  bot filter, supervision
              └──┬─────────┬────────┬──┘
                 │         │        │      loopback only
            ┌────▼───┐ ┌───▼────┐ ┌─▼──────┐
            │ site A │ │ site B │ │ site C │  own process, own
            └────────┘ └────────┘ └────────┘  plugins, config, logs

The front-end owns the public ports. It terminates TLS with each site's own certificate, decides which site a request belongs to — by host name, by the listener it arrived on, even by the network interface — and passes it to that site's own worker process.

  • Isolation by process. Each site runs in a separate address space with its own configuration, plugins, document root and logs. One site's bug cannot read another site's memory.
  • Every request is routed again. Browsers reuse connections. FFS does not assume the second request on a connection is for the same site as the first — and refuses a request for another site over a TLS connection that carries the first site's certificate.
  • The real visitor address, safely. Sites see the visitor's real address, because the guard bans by address. The forwarding header that carries it is accepted only from the front-end itself, so nobody can forge an address to get someone else banned.
  • Failure stays local. A site whose certificate is missing, whose port is taken or whose worker will not start is marked failed and answered with a maintenance status. The front-end restarts a worker that crashed, with growing back-off, and stops retrying one that never came up — an expired certificate does not improve with waiting.
  • Strangers get nothing. What a request for a host name no site claims receives is configurable — a page, a status, or a closed connection without a byte.

Groups: share what should be shared

Sites that belong together can form a group. A group member uses its own file where it has one and the group's where it does not: themes, images, translations, the contact form's settings, the guard's shared pattern lists and ban lists. Ten sites with one look and one mailbox are maintained in one place — and any site can still override any file.

Federation: a family of sites that feels like one

A group can present itself as a federation:

  • a site switcher in the navigation bar and the footer of every member, configured once for the group, translated per language;
  • a shared language choice — a visitor who switches to German on one site arrives in German on the next — or a separate choice per site, as the group decides.

The fedem sites — fedem.eu and the product and library sites next to it — are one federation on one FFS server.

Configured in one file

The whole server — listeners, sites, their host names and aliases, their certificates, groups — is one configuration file, checked before it is used:

ffs --check-config      report every problem, change nothing
ffs --dump-argv         show exactly how each site's worker will start
ffs --config ffs.fson   run