a route that exists only while something else agrees

Routing used to be a file per host, with the address of every backend written into it. That works until a service moves, and then the file is wrong in a way nothing reports. The migration replaced those entries with a directory that services register themselves into, and the result that mattered was that no address is written down any more: a service that is rescheduled stays reachable without an edit.

Then discovery failed in a way worth keeping. The directory publishes only the services whose health check passes, and every health check had been aimed at an address the service could not answer on from inside its own instance. So every service was unhealthy, the published set was empty, and the routing table was blank. A blank table looks exactly like a fleet with nothing running. The fix moved the checks inside the instance, asking the local port, and published the address the service actually listens on rather than the host's.

The second finding was ordering: the routes vanished across a restart of the routing process itself. So the migration shipped with the old file path retained as a fallback, and the directory now carries most of the routes while a handful stay where they were.

What it proves is that the lookup was the easy half. A directory is a claim about health, and the health signal has to be engineered as carefully as the address book it replaces.