Health-check the model, not the process

The practice

A service can be active while its model is still loading, and a metadata API can answer 200 with a dead backend. Probe something that actually runs inference, and expect one to two minutes of 503s after a restart.

2026-10-09