As reported by Mike Heffner on the mailing list:
We just tried rolling out an upgrade where we updated from the 2.0.6 driver to the 2.0.8 version. Immediately after pushing out the new version we noticed the [native protocol] open connections to the C* rings began climbing.
Upon further investigation, the connections keep climbing until all connection pools reach their max capacity. This happens on their production server and staging server (with a lesser load). If the driver is completely idle, the connection count remains constant.
The only change in the application is the 2.0.6->2.0.8 upgrade. The only custom option is read timeout set to 6 seconds.
I am suspecting a bug in the idle connections timeout mechanism (which is the only major change in the pool between 2.0.6 and 2.0.8).
Investigating this issue showed problems with the implementation for JAVA-419, which will be reconsidered. Due to planning constraint, we cannot include that in 2.0.9, so I'm scheduling it for 2.0.10.
Closing this issue, since the revisited implementation for should fix the problem.
On second thought, keeping it open as "testing" in case can think of a specific integration test for this while testing 419.