TransmitData() calls currently take 120s to time out. This led to the failure of test_metrics_are_zero reported in
IMPALA-3327. As a workaround, "impala-server.backends.client-cache.clients-in-use" was removed from the list of metrics to check.
As a permanent solution we should be able to prevent connections getting into this 'zombie' mode if they do periodically return-and-check-for-cancel.