Centralized Logging -- Single Point of Failure?

[Nate Klingenstein]

Looking at running a large clustered IdP and considering centralizing the logging of those IdP's based on this:

https://wiki.shibboleth.net/confluence/display/SHIB2/IdPProdLogging#IdPProdLogging-CentralizedLogs

My hesitation is that I'm concerned that this introduces a single point of failure.  If logging is not centralized and e.g. logging for some reason goes haywire on one cluster member, then that node's disk fills and the node dies and falls out of the pool.

If logging is centralized, then if the server(s) to which logging is performed is down for whatever reason, does that take your entire cluster down?