Most test suites cannot test long-term stability. We once had a commercial J2EE app that would run fine for weeks on JDK 7.x but commit continual-GC suicide or burn 100% of all cores after about 24 hours on JDK 7.x+1.

After this happened with multiple oracle “bug fix” releases, we of course stopped upgrading as much as possible, spending days trying to determine if each security hole being patched in the JDK was actually an attack vector for this app.

We also implemented periodic “therapeutic restarts” on all app servers.

It was like running a critical server app on Windows 95.