Control connection is connected to "host4"
host4 gets removed from the cluster
later, the driver tries to use the control connection (for example, to refresh another host's info). This fails, and no reconnection attempt is scheduled.
The bug came from this check in ControlConnection#signalError:
The assumption that a reconnect was already triggered is not true for removed hosts. The code used to work around that by checking host metadata, but this was accidentally lost in commit 04479ae.
Reestablished the metadata check and also added an explicit reconnect in ControlConnection#onRemove, which seems better than waiting for the connection to fail.
Created integration test (https://github.com/datastax/java-driver/pull/260) and validated against 2.0 and 2.1. The test scenario does the following:
Create 3 node CCM Cluster.
Create and initialize a Cluster instance with a single contact point of node1.
Ensures that the current control connection is connected to node1.
Ensures that the control connection is connected to another node.