I did some reading on how Cassandra does its internal checking and it implements a phi-accrual detection algorithm which is more sensitive to dynamic network conditions than a simple heartbeat. It also provides a scalar failure measurement instead of a binary yes/no detection which allows for configuration of tolerance levels.

See:
http://ddg.jaist.ac.jp/pub/HDY+04.pdf

There are pros/cons (particularly around simplicity), but I'd be curious what you at 10gen think about the appropriateness/usefulness of basing your failure detection off of this kind of a protocol.