When would you write a custom HealthIndicator for a Spring Boot service?

I define the consequence before the implementation: will this result remove one instance, page an operator, or only enrich diagnosis? Checks are isolated from request pools where possible, resistant to cascades, and may use a short-lived cached result if polling would overload the dependency. Status details never include credentials, connection strings, tenant data, or raw exception bodies. Failure and recovery behavior are tested under partial outages, because a noisy or sticky custom check can be more damaging than no check.