Re:Scary analogy (Score 1)

When you restart the master server for the production database, what will the application servers query?

The next node in the cluster? If this is a critical infrastructure system where time is money (or lives, or whatever) then you DO have a redundant highly available architecture, right?
 

The fact that "all systems develop runtime cruft over time" is the problem.

This is only part of the problem (and I'm not convinced it is fact). Other problems are that hardware has a finite lifetime (be it however long), and we never have complete control over the environment, no matter how much we spend on the space.

No matter how how bulletproof the system is, nothing is completely immune to failure - I do, however, rebut the notion that rebooting and kernel updates are inevitably linked. Something /will/ crash, and murphy dictates that this will not happen when you plan it - so it is best to plan accordingly. This is a trait of a good sysadmin. This makes a reboot after a kernel upgrade a sensible precaution - but not inevitable, and perhaps the reboot is better aligned with a disaster recovery planning operation.

Ultimately, as previously noted, the major advantage to this is not skipping an otherwise obligatory reboot, rather, it is being able to segregate upgrades and reboots, and not be forced to combine those activities. Both are important, but more control of potential downtime activities is a good thing.