If the database goes away, restarting every application instance is unlikely to bring it back. Yet that is exactly the recovery plan a health endpoint can accidentally create. Put every dependency behind one green-or-red response, connect it to the restart mechanism and a shared outage becomes a collection of restarting processes.
I like the distinction in the Kubernetes probe documentation. Liveness helps decide whether a container needs restarting. Readiness determines whether it should receive traffic from matching Services. A startup probe gives initialization its own allowance before the other probes run. Those checks can share implementation where that makes sense, but the decisions they drive are different.
For liveness, I would ask what failure restarting this process can actually repair. A stuck application may qualify. An unavailable external API usually needs a different response. For readiness, check what the instance needs to serve the traffic it is about to receive. If it can still handle useful requests while an optional integration is down, marking the entire instance unavailable may throw away working functionality. If the database is essential to every request, pretending the instance is ready is not helpful either.
Keep the checks cheap and bounded, and test the actions they cause. Take a required dependency away in an isolated environment and watch whether the application stops receiving traffic, restarts or continues in a useful degraded state. Check recovery as well as failure. The important result is that traffic and restarts behave sensibly; the health endpoint returning the expected colour is only part of that.