Inheriting a shared dependency's failures
Federating authentication out to a shared identity provider concentrates a real single point of failure: every relying application now depends on that one provider being reachable, and an outage there potentially affects every one of them simultaneously rather than just one isolated service. This is usually still the right tradeoff — the alternative of each application managing its own authentication independently carries its own, generally worse, risks — but it means an identity provider's reliability deserves the same scrutiny any other single point of failure would.
Graceful degradation is a relying application's deliberately designed, reduced-capability behavior for exactly this scenario: continuing to serve what it safely can when the identity provider is unreachable, rather than failing completely the moment that dependency goes down.
What 'graceful' actually looks like here
A common, reasonable pattern lets already-authenticated sessions continue working uninterrupted while new logins are blocked until the provider recovers — someone mid-session doesn't lose their work because of an outage that has nothing to do with what they're currently doing, while a new login attempt correctly can't be authenticated at all without the provider actually being reachable to authenticate it. This has to be designed and tested deliberately in advance; without that groundwork, what actually happens during a real outage is whatever the application does by default, which is rarely the graceful behavior anyone would have chosen on purpose.