One feature that didn't make it into the massive Cassandra 2.0 release was the rapid read protection feature, but it's here now in this week's 2.0.2 release. According to Cassandra frontman Johnathan Ellis, "Rapid read protection allows Cassandra to tolerate node failure without dropping a single request."
There's a great illustration on the blog post that shows a huge dip in performance when a node is killed 'with prejudice.' Cassandra brings traffic to a halt in this scenario until the dead node is caught by failure detection and taken out of service. Rapid read protection, which seems to be enabled by default in Cassandra 2.0.2 will prevent that.